Andreas Geiger (MPI Tübingen) | Philip Lenz (KIT) | Christoph Stiller (KIT) | Raquel Urtasun (University of Toronto)

Object Detection Evaluation 2012


The object detection and object orientation estimation benchmark consists of 7481 training images and 7518 test images, comprising a total of 80.256 labeled objects. All images are color and saved as png. For evaluation, we compute precision-recall curves for object detection and orientation-similarity-recall curves for joint object detection and orientation estimation. In the latter case not only the object 2D bounding box has to be located correctly, but also the orientation estimate in bird's eye view is evaluated. To rank the methods we compute average precision and average orientation similiarity. We require that all methods use the same parameter set for all test pairs. Our development kit provides details about the data format as well as MATLAB / C++ utility functions for reading and writing the label files.

We evaluate object detection performance using the PASCAL criteria and object detection and orientation estimation performance using the measure discussed in our CVPR 2012 publication. For cars we require an overlap of 70%, while for pedestrians and cyclists we require an overlap of 50% for a detection. Detections in don't care areas or detections which are smaller than the minimum size do not count as false positive. Difficulties are defined as follows:

  • Easy: Min. bounding box height: 40 Px, Max. occlusion level: Fully visible, Max. truncation: 15 %
  • Moderate: Min. bounding box height: 25 Px, Max. occlusion level: Partly occluded, Max. truncation: 30 %
  • Hard: Min. bounding box height: 25 Px, Max. occlusion level: Difficult to see, Max. truncation: 50 %

All methods are ranked based on the moderately difficult results. Note that for the hard evaluation ~2 % of the provided bounding boxes have not been recognized by humans, thereby upper bounding recall at 98 %. Hence, the hard evaluation is only given for reference.
Note 1: On 25.04.2017, we have fixed a bug in the object detection evaluation script. As of now, the submitted detections are filtered based on the min. bounding box height for the respective category which we have been done before only for the ground truth detections, thus leading to false positives for the category "Easy" when bounding boxes of height 25-39 Px were submitted (and to false positives for all categories if bounding boxes smaller than 25 Px were submitted). We like to thank Amy Wu, Matt Wilder, Pekka Jänis and Philippe Vandermersch for their feedback. The last leaderboards right before the changes can be found here!

Important Policy Update: As more and more non-published work and re-implementations of existing work is submitted to KITTI, we have established a new policy: from now on, only submissions with significant novelty that are leading to a peer-reviewed paper in a conference or journal are allowed. Minor modifications of existing algorithms or student research projects are not allowed. Such work must be evaluated on a split of the training set. To ensure that our policy is adopted, new users must detail their status, describe their work and specify the targeted venue during registration. Furthermore, we will regularly delete all entries that are 6 months old but are still anonymous or do not have a paper associated with them. For conferences, 6 month is enough to determine if a paper has been accepted and to add the bibliography information. For longer review cycles, you need to resubmit your results.
Additional information used by the methods
  • Stereo: Method uses left and right (stereo) images
  • Flow: Method uses optical flow (2 temporally adjacent images)
  • Multiview: Method uses more than 2 temporally adjacent images
  • Laser Points: Method uses point clouds from Velodyne laser scanner
  • Additional training data: Use of additional data sources for training (see details)

Car


Method Setting Code Moderate Easy Hard Runtime Environment
1 THU CV-AI 91.97 % 91.96 % 84.57 % 0.38 s GPU @ 2.5 Ghz (Python)
2 DH-ARI 91.48 % 90.87 % 82.25 % 4s GPU @ 2.5 Ghz (C/C++)
3 HRI-SH 90.71 % 91.34 % 84.28 % 3.6 s GPU @ >3.5 Ghz (Python + C/C++)
4 BM-NET 90.50 % 90.81 % 83.92 % 0.5 s GPU @ 2.5 Ghz (Python + C/C++)
5 MVRA + I-FRCNN+ 90.36 % 90.78 % 80.48 % 0.18 s GPU @ 2.5 Ghz (Python)
6 TuSimple code 90.33 % 90.77 % 82.86 % 1.6 s GPU @ 2.5 Ghz (Python + C/C++)
F. Yang, W. Choi and Y. Lin: Exploit all the layers: Fast and accurate cnn object detector with scale dependent pooling and cascaded rejection classifiers. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016.
K. He, X. Zhang, S. Ren and J. Sun: Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition 2016.
7 RRC code 90.23 % 90.61 % 87.44 % 3.6 s GPU @ 2.5 Ghz (C/C++)
J. Ren, X. Chen, J. Liu, W. Sun, J. Pang, Q. Yan, Y. Tai and L. Xu: Accurate Single Stage Detector Using Recurrent Rolling Convolution. CVPR 2017.
8 CFENet 90.22 % 90.33 % 84.85 % 4 s GPU @ 2.5 Ghz (Python + C/C++)
9 UberATG-MMF
This method makes use of Velodyne laser scans.
 90.17 % 91.82 % 88.54 % 0.08 s GPU @ 2.5 Ghz (Python)
M. Liang*, B. Yang*, Y. Chen, R. Hu and R. Urtasun: Multi-Task Multi-Sensor Fusion for 3D Object Detection. CVPR 2019.
10 PC-CNN-V2
This method makes use of Velodyne laser scans.
 90.15 % 90.79 % 87.58 % 0.5 s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Du, M. Ang, S. Karaman and D. Rus: A General Pipeline for 3D Detection of Vehicles. 2018 IEEE International Conference on Robotics and Automation (ICRA) 2018.
11 EM-FPS 90.15 % 90.61 % 84.01 % 0.15 s GPU @ 1.5 Ghz (Python + C/C++)
12 Fusion-RCNN 90.10 % 90.75 % 81.09 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
13 THICV-YDM 90.08 % 90.40 % 80.41 % 0.06 s GPU @ 2.5 Ghz (Python)
14 SJTU-HW 90.08 % 90.81 % 79.98 % 0.85s GPU @ 1.5 Ghz (Python + C/C++)
S. Zhang, X. Zhao, L. Fang, F. Haiping and S. Haitao: LED: LOCALIZATION-QUALITY ESTIMATION EMBEDDED DETECTOR. IEEE International Conference on Image Processing 2018.
L. Fang, X. Zhao and S. Zhang: Small-objectness sensitive detection based on shifted single shot detector. Multimedia Tools and Applications 2018.
15 MDC
This method makes use of Velodyne laser scans.
 90.03 % 90.72 % 80.87 % 0.17 s GPU @ 2.5 Ghz (Python)
16 Deep MANTA 90.03 % 97.25 % 80.62 % 0.7 s GPU @ 2.5 Ghz (Python + C/C++)
F. Chabot, M. Chaouch, J. Rabarisoa, C. Teulière and T. Chateau: Deep MANTA: A Coarse-to-fine Many-Task Network for joint 2D and 3D vehicle analysis from monocular image. CVPR 2017.
17 sensekitti code 90.00 % 90.76 % 81.83 % 4.5 s GPU @ 2.5 Ghz (Python + C/C++)
B. Yang, J. Yan, Z. Lei and S. Li: Craft Objects from Images. CVPR 2016.
18 F-PointNet
This method makes use of Velodyne laser scans.
 code 90.00 % 90.78 % 80.80 % 0.17 s GPU @ 3.0 Ghz (Python)
C. Qi, W. Liu, C. Wu, H. Su and L. Guibas: Frustum PointNets for 3D Object Detection from RGB-D Data. arXiv preprint arXiv:1711.08488 2017.
19 ECV-NET 89.93 % 90.61 % 81.81 % 0.4 s GPU @ 2.5 Ghz (C/C++)
20 ART-Det 89.89 % 95.18 % 80.03 % 0.067s GPU @ 2.5 Ghz (Python + C/C++)
21 FNV2 89.88 % 90.51 % 80.66 % 0.18 s GPU @ 2.5 Ghz (Python)
22 FichaDL 89.86 % 95.79 % 80.27 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
23 Det-RGBD
This method uses stereo information.
 89.83 % 90.45 % 80.56 % 0.58 s GPU @ 2.5 Ghz (Python + C/C++)
24 MBR-SSD 89.82 % 90.32 % 82.28 % 4.0 s GPU @ 2.5 Ghz (Python + C/C++)
25 F-ConvNet
This method makes use of Velodyne laser scans.
 89.79 % 90.44 % 80.66 % 0.47 s GPU @ 2.5 Ghz (Python + C/C++)
Z. Wang and K. Jia: Frustum ConvNet: Sliding Frustums to Aggregate Local Point-Wise Features for Amodal 3D Object Detection. arXiv preprint arXiv:1903.01864 2019.
26 PointRCNN-deprecated
This method makes use of Velodyne laser scans.
 89.75 % 90.77 % 80.98 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
27 SINet+ code 89.73 % 90.51 % 77.82 % 0.3 s TITAN X GPU
X. Hu, X. Xu, Y. Xiao, H. Chen, S. He, J. Qin and P. Heng: SINet: A Scale-insensitive Convolutional Neural Network for Fast Vehicle Detection. IEEE Transactions on Intelligent Transportation Systems 2019.
28 STD 89.72 % 90.57 % 88.90 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
29 Part-A^2
This method makes use of Velodyne laser scans.
 89.71 % 90.69 % 88.41 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
30 Fast Point R-CNNv1.1
This method makes use of Velodyne laser scans.
 89.71 % 90.59 % 88.13 % 0.06 s GPU @ 2.5 Ghz (Python + C/C++)
31 Alibaba-AILabsX
This method makes use of Velodyne laser scans.
 89.69 % 90.59 % 80.83 % 0.2 s GPU @ >3.5 Ghz (Python)
32 AILabs3D
This method makes use of Velodyne laser scans.
 89.68 % 90.57 % 80.67 % 0.6 s GPU @ >3.5 Ghz (Python)
33 Alibaba-AILabsX
This method makes use of Velodyne laser scans.
 89.65 % 90.47 % 80.81 % 0.05 s 1 core @ >3.5 Ghz (C/C++)
34 Aston-EAS 89.64 % 90.49 % 77.95 % 0.24 s GPU @ 2.5 Ghz (Python + C/C++)
J. Wei, J. He, Y. Zhou, K. Chen, Z. Tang and Z. Xiong: Enhanced Object Detection With Deep Convolutional Neural Networks for Advanced Driving Assistance. IEEE Transactions on Intelligent Transportation Systems 2019.
35 Patches
This method makes use of Velodyne laser scans.
 89.61 % 90.75 % 87.42 % 0.15 s GPU @ 2.0 Ghz
36 Voxel-FPN 89.60 % 90.51 % 80.78 % 0.02 s GPU @ 2.5 Ghz (Python + C/C++)
37 epBRM
This method makes use of Velodyne laser scans.
 89.58 % 90.50 % 87.51 % 0.1 s GPU @ >3.5 Ghz (Python + C/C++)
38 SINet_VGG code 89.56 % 90.60 % 78.19 % 0.2 s TITAN X GPU
X. Hu, X. Xu, Y. Xiao, H. Chen, S. He, J. Qin and P. Heng: SINet: A Scale-insensitive Convolutional Neural Network for Fast Vehicle Detection. IEEE Transactions on Intelligent Transportation Systems 2019.
39 Patches - EMP
This method makes use of Velodyne laser scans.
 89.55 % 94.64 % 88.02 % 0.5 s GPU @ 2.5 Ghz (Python)
40 PFPN 89.54 % 90.52 % 80.76 % 0.02 s 4 cores @ >3.5 Ghz (Python)
41 Fast Point R-CNN
This method makes use of Velodyne laser scans.
 89.51 % 90.58 % 87.89 % 0.06 s GPU @ 2.5 Ghz (Python + C/C++)
42 DH-ARI 89.47 % 90.31 % 84.78 % 0.2 s 1 core @ >3.5 Ghz (Python + C/C++)
43 SDP+RPN 89.42 % 89.90 % 78.54 % 0.4 s GPU @ 2.5 Ghz (Python + C/C++)
F. Yang, W. Choi and Y. Lin: Exploit All the Layers: Fast and Accurate CNN Object Detector with Scale Dependent Pooling and Cascaded Rejection Classifiers. Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition 2016.
S. Ren, K. He, R. Girshick and J. Sun: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 2015.
44 VAT-Net 89.41 % 90.69 % 79.97 % 1 s GPU @ 2.5 Ghz (Python + C/C++)
45 SegVoxelNet 89.37 % 90.62 % 88.02 % 0.04 s 1 core @ 2.5 Ghz (Python)
46 ZRNet(ResNet-50) 89.34 % 89.91 % 79.28 % 0.04 s GPU @ 2.5 Ghz (Python)
47 PointRCNN
This method makes use of Velodyne laser scans.
 code 89.32 % 90.74 % 85.73 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
S. Shi, X. Wang and H. Li: PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud. CVPR 2019.
48 IPOD 89.30 % 90.20 % 87.37 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
49 TBA 89.30 % 90.42 % 87.17 % 0.07 s 1 core @ 2.5 Ghz (Python)
50 RAL
This method makes use of Velodyne laser scans.
 89.26 % 90.34 % 80.62 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
51 MVX-Net
This method makes use of Velodyne laser scans.
 89.25 % 90.43 % 80.47 % 0.06 s GPU @ 3.0 Ghz (Python + C/C++)
52 ITVD code 89.23 % 90.57 % 79.31 % 0.3 s GPU @ 2.5 Ghz (C/C++)
Y. Wei Liu: Improving Tiny Vehicle Detection in Complex Scenes. IEEE International Conference on Multimedia and Expo (ICME) 2018.
53 PointPillars
This method makes use of Velodyne laser scans.
 code 89.22 % 90.33 % 87.04 % 16 ms 1080ti GPU and Intel i7 CPU
A. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang and O. Beijbom: PointPillars: Fast Encoders for Object Detection from Point Clouds. CVPR 2019.
54 MMV 89.22 % 90.59 % 80.49 % 0.4 s GPU @ 2.5 Ghz (C/C++)
55 MPNet
This method makes use of Velodyne laser scans.
 89.21 % 90.60 % 86.19 % 0.02 s GPU @ 2.5 Ghz (Python + C/C++)
56 CONV-BOX
This method makes use of Velodyne laser scans.
 89.20 % 90.35 % 87.88 % 0.2 s Tesla V100
57 VCTNet 89.20 % 89.60 % 80.04 % 0.02 s GPU @ 1.5 Ghz (C/C++)
58 4D-MSCNN+CRL
This method uses stereo information.
 89.19 % 90.32 % 76.26 % 0.2 s GPU @ 2.5 Ghz (Matlab + C/C++)
59 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 89.17 % 90.43 % 85.82 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
60 Sogo_MM 89.17 % 90.80 % 79.58 % 1.5 s GPU @ 2.5 Ghz (C/C++)
61 MV3D
This method makes use of Velodyne laser scans.
 89.17 % 90.53 % 80.16 % 0.36 s GPU @ 2.5 Ghz (Python + C/C++)
X. Chen, H. Ma, J. Wan, B. Li and T. Xia: Multi-View 3D Object Detection Network for Autonomous Driving. CVPR 2017.
62 NU-optim 89.17 % 90.24 % 87.92 % 0.04 s GPU @ >3.5 Ghz (Python)
63 A-VoxelNet 89.15 % 90.27 % 80.43 % 0.029 s GPU @ 2.5 Ghz (Python)
64 MVSLN 89.14 % 90.65 % 80.64 % 0.1s s 1 core @ 2.5 Ghz (C/C++)
65 SRF 89.14 % 90.34 % 80.52 % 0.05 s GPU @ 2.5 Ghz (Python + C/C++)
66 YOLOv3.5 89.10 % 89.70 % 79.79 % 0.05 s GPU @ 2.5 Ghz (Python)
67 SINet_PVA code 89.08 % 90.44 % 75.85 % 0.11 s TITAN X GPU
X. Hu, X. Xu, Y. Xiao, H. Chen, S. He, J. Qin and P. Heng: SINet: A Scale-insensitive Convolutional Neural Network for Fast Vehicle Detection. IEEE Transactions on Intelligent Transportation Systems 2019.
68 SECOND-V1.5
This method makes use of Velodyne laser scans.
 code 89.06 % 90.52 % 80.40 % 0.04 s GPU @ 2.0 Ghz (Python + C/C++)
69 AtrousDet 89.01 % 90.25 % 78.98 % 0.05 s TITAN X
70 CLA 88.99 % 90.51 % 75.50 % 0.3 s GPU @ 2.5 Ghz (Matlab + C/C++)
71 InNet 88.95 % 90.26 % 79.46 % 0.16 s GPU @ 3.5 Ghz (Python + C/C++)
72 Shift R-CNN 88.90 % 90.56 % 79.86 % 0.25 s GPU @ 1.5 Ghz (Python)
A. Naiden, V. Paunescu, G. Kim, B. Jeon and M. Leordeanu: Shift R-CNN: Deep Monocular 3D Object Detection With Closed-form Geometric Constraints. ICIP 2019.
73 SubCNN 88.86 % 90.75 % 79.24 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
Y. Xiang, W. Choi, Y. Lin and S. Savarese: Subcategory-aware Convolutional Neural Networks for Object Proposals and Detection. IEEE Winter Conference on Applications of Computer Vision (WACV) 2017.
74 Deep3DBox 88.86 % 90.47 % 77.60 % 1.5 s GPU @ 2.5 Ghz (C/C++)
A. Mousavian, D. Anguelov, J. Flynn and J. Kosecka: 3D Bounding Box Estimation Using Deep Learning and Geometry. CVPR 2017.
75 MonoPSR 88.84 % 90.18 % 71.44 % 0.2 s GPU @ 3.5 Ghz (Python)
J. Ku*, A. Pon* and S. Waslander: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. CVPR 2019.
76 FQNet 88.83 % 90.45 % 77.55 % 0.5 s 1 core @ 2.5 Ghz (Python)
77 MS-CNN code 88.83 % 90.46 % 74.76 % 0.4 s GPU @ 2.5 Ghz (C/C++)
Z. Cai, Q. Fan, R. Feris and N. Vasconcelos: A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection. ECCV 2016.
78 RAR-Net 88.83 % 90.45 % 77.55 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
79 ZRNet 88.82 % 89.77 % 79.07 % 0.04 s GPU @ 2.5 Ghz (Python)
80 CFR
This method makes use of Velodyne laser scans.
 88.77 % 90.53 % 80.23 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
81 RCN-resnet101 88.75 % 89.08 % 79.97 % 0.3 s GPU @ 2.5 Ghz (Python + C/C++)
82 DeepStereoOP 88.75 % 90.34 % 79.39 % 3.4 s GPU @ 3.5 Ghz (Matlab + C/C++)
C. Pham and J. Jeon: Robust Object Proposals Re-ranking for Object Detection in Autonomous Driving Using Convolutional Neural Networks. Signal Processing: Image Communiation 2017.
83 3DBN
This method makes use of Velodyne laser scans.
 88.62 % 90.30 % 80.08 % 0.13s 1080Ti (Python+C/C++)
X. Li, J. Guivant, N. Kwok and Y. Xu: 3D Backbone Network for 3D Object Detection. CoRR 2019.
84 SAG-Net 88.61 % 89.25 % 79.72 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
85 PP_v1.0 code 88.57 % 90.41 % 84.23 % 0.02s 1 core @ 2.5 Ghz (C/C++)
86 PAD 88.47 % 90.36 % 83.19 % 0.15 s 1 core @ 2.5 Ghz (Python)
87 ARPNET 88.44 % 90.26 % 79.86 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
88 SECOND code 88.40 % 90.40 % 80.21 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
89 3DOP
This method uses stereo information.
 code 88.34 % 90.09 % 78.79 % 3s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Y. Zhu, A. Berneshawi, H. Ma, S. Fidler and R. Urtasun: 3D Object Proposals for Accurate Object Class Detection. NIPS 2015.
90 DA 88.23 % 90.45 % 74.21 % 0.08 s 1 core @ 2.5 Ghz (C/C++)
91 MM-MRFC
This method uses optical flow information.
This method makes use of Velodyne laser scans.
 88.20 % 90.93 % 78.02 % 0.05 s GPU @ 2.5 Ghz (C/C++)
A. Costea, R. Varga and S. Nedevschi: Fast Boosting based Detection using Scale Invariant Multimodal Multiresolution Filtered Features. CVPR 2017.
92 AVOD
This method makes use of Velodyne laser scans.
 code 88.08 % 89.73 % 80.14 % 0.08 s Titan X (pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
93 Multi-3D
This method makes use of Velodyne laser scans.
 88.04 % 90.10 % 75.31 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
94 cas+res+soft 88.00 % 89.82 % 77.66 % 0.2 s 4 cores @ 2.5 Ghz (Python)
95 SeoulRobotics-HFD
This method makes use of Velodyne laser scans.
 87.86 % 90.02 % 79.95 % 0.035 s GPU (C++)
96 Mono3D code 87.86 % 90.27 % 78.09 % 4.2 s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Z. Zhang, H. Ma, S. Fidler and R. Urtasun: Monocular 3D Object Detection for Autonomous Driving. CVPR 2016.
97 merge12-12 87.81 % 89.88 % 77.42 % 0.2 s 4 cores @ 2.5 Ghz (Python)
98 DFD 87.78 % 90.02 % 79.78 % 0.05 s GPU @ 2.0 Ghz (Python + C/C++)
99 MonoDIS 87.58 % 90.31 % 76.85 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
100 TridentNet 87.49 % 88.38 % 78.97 % 0.2 s GPU @ 2.5 Ghz (Python)
101 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 87.44 % 89.99 % 80.05 % 0.1 s Titan X (Pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
102 SECA 87.42 % 89.57 % 79.43 % 0.09 s GPU @ 2.5 Ghz (Python)
103 SCANet 87.31 % 89.34 % 79.30 % 0.09s GPU @ 2.5 Ghz (Python)
104 SCANet 87.12 % 89.65 % 79.43 % 0.17 s >8 cores @ 2.5 Ghz (Python)
105 ODES code 87.10 % 86.82 % 78.32 % 0.02 s GPU @ 2.5 Ghz (Python)
ERROR: Wrong syntax in BIBTEX file.
106 ELLIOT
This method makes use of Velodyne laser scans.
 86.98 % 90.20 % 81.41 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
107 cas_retina 86.23 % 89.53 % 75.77 % 0.2 s 4 cores @ 2.5 Ghz (Python)
108 cascadercnn 85.86 % 84.21 % 69.57 % 0.36 s 4 cores @ 2.5 Ghz (Python)
109 ReSqueeze 85.74 % 87.12 % 77.02 % 0.03 s GPU @ >3.5 Ghz (Python)
110 NLK 85.56 % 89.00 % 79.34 % 0.02 s 1 core @ 2.5 Ghz (Python)
111 monocular 85.42 % 87.33 % 77.43 % 0.4 s GPU @ 2.5 Ghz (Python + C/C++)
112 anm 85.33 % 90.11 % 76.55 % 3 s 1 core @ 2.5 Ghz (C/C++)
113 IoU_DCRCNN 84.48 % 87.68 % 76.70 % 0.66 s GPU @ 2.5 Ghz (Python)
114 cas_retina_1_13 84.43 % 89.22 % 75.39 % 0.03 s 4 cores @ 2.5 Ghz (Python)
115 YOLOv3+d 84.13 % 84.30 % 76.34 % 0.04 s GPU @ 1.5 Ghz (C/C++)
116 StereoFENet
This method uses stereo information.
 83.65 % 89.01 % 77.12 % 0.15 s 1 core @ 3.5 Ghz (Python)
W. Bao, B. Xu and Z. Chen: MonoFENet: Monocular 3D Object Detection with Feature Enhancement Networks. 2019.
117 NEUAV 83.25 % 87.75 % 76.38 % 0.06 s GPU @ 2.5 Ghz (Python)
118 MonoFENet 82.54 % 89.10 % 76.39 % 0.15 s 1 core @ 3.5 Ghz (Python)
W. Bao, B. Xu and Z. Chen: MonoFENet: Monocular 3D Object Detection with Feature Enhancement Networks. 2019.
119 cascade_gw 82.24 % 81.09 % 67.95 % 0.2 s 4 cores @ 2.5 Ghz (Python)
120 LPN 81.67 % 87.70 % 72.69 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
121 A3DODWTDA (image) code 81.54 % 76.21 % 66.85 % 0.8 s GPU @ 3.0 Ghz (Python)
F. Gustafsson and E. Linder-Norén: Automotive 3D Object Detection Without Target Domain Annotations. 2018.
122 SDP+CRC (ft) 81.33 % 90.39 % 70.33 % 0.6 s GPU @ 2.5 Ghz (C/C++)
F. Yang, W. Choi and Y. Lin: Exploit All the Layers: Fast and Accurate CNN Object Detector with Scale Dependent Pooling and Cascaded Rejection Classifiers. Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition 2016.
123 ResNet-RRC w/RGBD 81.09 % 89.91 % 71.78 % 0.057 s GPU @ 1.5 Ghz (Python + C/C++)
124 ResNet-RRC (Adv. HW) 81.00 % 89.89 % 71.56 % 0.06 s GPU @ 1.5 Ghz (Python + C/C++)
125 Stereo R-CNN
This method uses stereo information.
 code 80.80 % 90.23 % 71.42 % 0.3 s GPU @ 2.5 Ghz (Python)
P. Li, X. Chen and S. Shen: Stereo R-CNN based 3D Object Detection for Autonomous Driving. CVPR 2019.
126 X_MD 80.65 % 89.81 % 79.66 % 0.2 s 1 core @ 2.5 Ghz (Python + C/C++)
127 FNV1_Fusion 80.41 % 89.37 % 79.03 % 0.11 s GPU @ 2.5 Ghz (Python)
128 FNV1_RPN 80.41 % 89.44 % 79.14 % 0.12 s 1 core @ 2.5 Ghz (Python + C/C++)
129 Cmerge 80.25 % 89.83 % 70.76 % 0.2 s 4 cores @ 2.5 Ghz (Python)
130 SS3D 80.11 % 89.15 % 70.52 % 48 ms Tesla V100 (Python)
131 SECA 80.05 % 89.26 % 78.80 % 1 s GPU @ 2.5 Ghz (Python)
132 VSE 80.05 % 89.26 % 78.80 % 0.15 s GPU @ 2.5 Ghz (Python)
133 BS3D 80.02 % 89.85 % 70.14 % 22 ms Titan Xp
134 MV3D (LIDAR)
This method makes use of Velodyne laser scans.
 79.76 % 89.80 % 78.61 % 0.24 s GPU @ 2.5 Ghz (Python + C/C++)
X. Chen, H. Ma, J. Wan, B. Li and T. Xia: Multi-View 3D Object Detection Network for Autonomous Driving. CVPR 2017.
135 ZKNet 79.62 % 89.89 % 70.03 % 0.01 s GPU @ 2.0 Ghz (Python)
136 Pseudo-LiDAR V2
This method uses stereo information.
 79.59 % 90.18 % 76.20 % 0.4 s GPU @ 2.5 Ghz (Python)
137 RFCN_RFB 79.45 % 83.85 % 67.51 % 0.2 s 4 cores @ 2.5 Ghz (Python)
138 Complexer-YOLO
This method makes use of Velodyne laser scans.
 79.31 % 88.11 % 79.11 % 0.06 s GPU @ 3.5 Ghz (C/C++)
139 FNV1 79.28 % 88.45 % 77.14 % 0.11 s GPU @ 2.5 Ghz (Python)
140 RefineNet 79.21 % 90.16 % 65.71 % 0.20 s GPU @ 2.5 Ghz (Matlab + C++)
R. Rajaram, E. Bar and M. Trivedi: RefineNet: Refining Object Detectors for Autonomous Driving. IEEE Transactions on Intelligent Vehicles 2016.
R. Rajaram, E. Bar and M. Trivedi: RefineNet: Iterative Refinement for Accurate Object Localization. Intelligent Transportation Systems Conference 2016.
141 softretina 79.15 % 89.36 % 69.24 % 0.16 s 4 cores @ 2.5 Ghz (Python)
142 Faster R-CNN code 79.11 % 87.90 % 70.19 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
S. Ren, K. He, R. Girshick and J. Sun: Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks. NIPS 2015.
143 detectron code 78.96 % 88.14 % 69.74 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
144 FRCNN+Or code 78.95 % 89.87 % 68.97 % 0.09 s Titan Xp GPU
C. Guindel, D. Martin and J. Armingol: Fast Joint Object Detection and Viewpoint Estimation for Traffic Scene Understanding. IEEE Intelligent Transportation Systems Magazine 2018.
C. Guindel, D. Martin and J. Armingol: Joint Object Detection and Viewpoint Estimation using CNN features. IEEE International Conference on Vehicular Electronics and Safety (ICVES) 2017.
145 Resnet101Faster rcnn 78.93 % 87.97 % 69.80 % 1 s 1 core @ 2.5 Ghz (Python)
146 Retinanet100 78.85 % 89.83 % 68.73 % 0.2 s 4 cores @ 2.5 Ghz (Python)
147 NM code 78.77 % 89.04 % 68.69 % 0.01 s GPU @ 2.5 Ghz (Python)
148 SeRC 78.33 % 88.28 % 69.36 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
149 Manhnet 78.03 % 85.86 % 61.13 % 26 ms 1 core @ 2.5 Ghz (C/C++)
150 RTL3D 77.63 % 76.95 % 71.17 % 0.02 s GPU @ 2.5 Ghz (Python)
151 avodC 77.54 % 86.86 % 70.00 % 0.1 s GPU @ 2.5 Ghz (Python)
152 MonoGRNet code 77.46 % 87.23 % 61.12 % 0.04s NVIDIA P40
Z. Qin, J. Wang and Y. Lu: MonoGRNet: A Geometric Reasoning Network for 3D Object Localization. The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19) 2019.
153 spLBP 77.39 % 80.16 % 60.59 % 1.5 s 8 cores @ 2.5 Ghz (Matlab + C/C++)
Q. Hu, S. Paisitkriangkrai, C. Shen, A. Hengel and F. Porikli: Fast Detection of Multiple Objects in Traffic Scenes With a Common Detection Framework. IEEE Trans. Intelligent Transportation Systems 2016.
154 SceneNet 77.34 % 87.90 % 68.38 % 0.03 s GPU @ 2.5 Ghz (C/C++)
155 CLF3D
This method makes use of Velodyne laser scans.
 77.00 % 84.51 % 67.81 % 0.13 s GPU @ 2.5 Ghz (Python)
156 MTDP 76.91 % 84.24 % 67.91 % 0.15 s GPU @ 2.0 Ghz (Python)
157 yolov3_warp 76.73 % 89.13 % 67.70 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
ERROR: Wrong syntax in BIBTEX file.
158 Reinspect code 76.65 % 88.36 % 66.56 % 2s 1 core @ 2.5 Ghz (C/C++)
R. Stewart, M. Andriluka and A. Ng: End-to-End People Detection in Crowded Scenes. CVPR 2016.
159 Regionlets 76.56 % 86.50 % 59.82 % 1 s >8 cores @ 2.5 Ghz (C/C++)
X. Wang, M. Yang, S. Zhu and Y. Lin: Regionlets for Generic Object Detection. T-PAMI 2015.
W. Zou, X. Wang, M. Sun and Y. Lin: Generic Object Detection with Dense Neural Patterns and Regionlets. British Machine Vision Conference 2014.
C. Long, X. Wang, G. Hua, M. Yang and Y. Lin: Accurate Object Detection with Location Relaxation and Regionlets Relocalization. Asian Conference on Computer Vision 2014.
160 AOG code 75.97 % 85.58 % 60.96 % 3 s 4 cores @ 2.5 Ghz (Matlab)
T. Wu, B. Li and S. Zhu: Learning And-Or Models to Represent Context and Occlusion for Car Detection and Viewpoint Estimation. TPAMI 2016.
B. Li, T. Wu and S. Zhu: Integrating Context and Occlusion for Car Detection by Hierarchical And-Or Model. ECCV 2014.
161 GS3D 75.84 % 83.92 % 60.24 % 2 s 1 core @ 2.5 Ghz (C/C++)
B. Li, W. Ouyang, L. Sheng, X. Zeng and X. Wang: GS3D: An Efficient 3D Object Detection Framework for Autonomous Driving. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2019.
162 3D FCN
This method makes use of Velodyne laser scans.
 75.83 % 85.54 % 68.30 % >5 s 1 core @ 2.5 Ghz (C/C++)
B. Li: 3D Fully Convolutional Network for Vehicle Detection in Point Cloud. IROS 2017.
163 3D-SSMFCNN code 75.78 % 75.51 % 67.75 % 0.1 s GPU @ 1.5 Ghz (C/C++)
L. Novak: Vehicle Detection and Pose Estimation for Autonomous Driving. 2017.
164 3DVP code 75.77 % 81.46 % 65.38 % 40 s 8 cores @ 3.5 Ghz (Matlab + C/C++)
Y. Xiang, W. Choi, Y. Lin and S. Savarese: Data-Driven 3D Voxel Patterns for Object Category Recognition. IEEE Conference on Computer Vision and Pattern Recognition 2015.
165 Pose-RCNN 75.74 % 88.89 % 61.86 % 2 s >8 cores @ 2.5 Ghz (Python)
M. Braun, Q. Rao, Y. Wang and F. Flohr: Pose-RCNN: Joint object detection and pose estimation using 3D object proposals. Intelligent Transportation Systems (ITSC), 2016 IEEE 19th International Conference on 2016.
166 SubCat code 75.46 % 81.45 % 59.71 % 0.7 s 6 cores @ 3.5 Ghz (Matlab + C/C++)
E. Ohn-Bar and M. Trivedi: Learning to Detect Vehicles by Clustering Appearance Patterns. T-ITS 2015.
167 myfaster-rcnn-v1.5 75.29 % 88.35 % 60.91 % 0.1 s 1 core @ 2.5 Ghz (Python)
168 multi-task CNN 75.21 % 83.45 % 66.89 % 25.1 ms GPU @ 2.0 Ghz (Python)
M. Oeljeklaus, F. Hoffmann and T. Bertram: A Fast Multi-Task CNN for Spatial Understanding of Traffic Scenes. IEEE Intelligent Transportation Systems Conference 2018.
169 A3DODWTDA
This method makes use of Velodyne laser scans.
 code 74.71 % 78.21 % 66.70 % 0.08 s GPU @ 3.0 Ghz (Python)
F. Gustafsson and E. Linder-Norén: Automotive 3D Object Detection Without Target Domain Annotations. 2018.
170 FD2 74.68 % 87.14 % 65.70 % 0.01 s GPU @ >3.5 Ghz (Python + C/C++)
171 BdCost+DA+MS 74.07 % 83.02 % 59.06 % TBD s 4 cores @ 2.5 Ghz (Matlab/C++)
172 RFCN 73.56 % 80.70 % 61.94 % 0.2 s 4 cores @ 2.5 Ghz (Python)
173 VoxelNet(Unofficial) 73.39 % 79.27 % 65.61 % 0.5 s GPU @ 2.0 Ghz (Python)
174 3DVSSD 73.39 % 84.39 % 65.64 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
175 bin 73.31 % 76.05 % 63.76 % 15ms s GPU @ >3.5 Ghz (Python)
176 yolo800 73.00 % 76.45 % 64.68 % 0.13 s 4 cores @ 2.5 Ghz (Python)
177 ResNet-RRC (Noised) 71.81 % 78.97 % 63.57 % .057 s GPU @ 1.5 Ghz (Python + C/C++)
178 Int-YOLO code 70.65 % 74.76 % 63.70 % 0.03 s 1 core @ 2.5 Ghz (C/C++)
ERROR: Wrong syntax in BIBTEX file.
179 MV-RGBD-RF
This method makes use of Velodyne laser scans.
 69.92 % 76.49 % 57.47 % 4 s 4 cores @ 2.5 Ghz (C/C++)
A. Gonzalez, D. Vazquez, A. Lopez and J. Amores: On-Board Object Detection: Multicue, Multimodal, and Multiview Random Forest of Local Experts.. IEEE Trans. on Cybernetics 2016.
A. Gonzalez, G. Villalonga, J. Xu, D. Vazquez, J. Amores and A. Lopez: Multiview Random Forest of Local Experts Combining RGB and LIDAR data for Pedestrian Detection. IEEE Intelligent Vehicles Symposium (IV) 2015.
180 AOG-View 69.89 % 84.29 % 57.25 % 3 s 1 core @ 2.5 Ghz (Matlab, C/C++)
B. Li, T. Wu and S. Zhu: Integrating Context and Occlusion for Car Detection by Hierarchical And-Or Model. ECCV 2014.
181 ROI-10D 69.64 % 75.33 % 61.18 % 0.2 s GPU @ 3.5 Ghz (Python)
F. Manhardt, W. Kehl and A. Gaidon: ROI-10D: Monocular Lifting of 2D Detection to 6D Pose and Metric Shape. Computer Vision and Pattern Recognition (CVPR) 2019.
182 fasterrcnn 68.76 % 73.64 % 59.72 % 0.2 s 4 cores @ 2.5 Ghz (Python)
183 MF3D 68.72 % 88.46 % 58.70 % 0.03 s GPU @ 2.5 Ghz (C/C++)
184 myfaster-rcnn 68.69 % 89.59 % 58.39 % 0.01 s 1 core @ 2.5 Ghz (Python)
185 Vote3Deep
This method makes use of Velodyne laser scans.
 68.39 % 76.95 % 63.22 % 1.5 s 4 cores @ 2.5 Ghz (C/C++)
M. Engelcke, D. Rao, D. Zeng Wang, C. Hay Tong and I. Posner: Vote3Deep: Fast Object Detection in 3D Point Clouds Using Efficient Convolutional Neural Networks. ArXiv e-prints 2016.
186 Pseudo-LiDAR
This method uses stereo information.
 code 67.96 % 85.08 % 59.55 % 0.4 s GPU @ 2.5 Ghz (C/C++)
Y. Wang, W. Chao, D. Garg, B. Hariharan, M. Campbell and K. Weinberger: Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving. CVPR 2019.
187 GPVL 67.89 % 77.76 % 58.23 % 10 s 1 core @ 2.5 Ghz (C/C++)
188 RFBnet 67.86 % 82.31 % 59.95 % 0.2 s 4 cores @ 2.5 Ghz (Python)
189 Fast-SSD 67.17 % 83.89 % 59.09 % 0.06 s GTX650Ti
190 BdCost48LDCF code 67.08 % 77.93 % 51.15 % 0.5 s 8 cores @ 3.5 Ghz (Matlab + C/C++)
A. Fernández-Baldera, J. Buenaposada and L. Baumela: BAdaCost: Multi-class Boosting with Costs . Pattern Recognition 2018.
191 SA_3D 66.69 % 86.36 % 55.18 % 0.3 s GPU @ 2.5 Ghz (Python)
192 OC-DPM 66.45 % 76.16 % 53.70 % 10 s 8 cores @ 2.5 Ghz (Matlab)
B. Pepik, M. Stark, P. Gehler and B. Schiele: Occlusion Patterns for Object Class Detection. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2013.
193 mymask-rcnn 66.31 % 86.32 % 54.33 % 0.3 s 1 core @ 2.5 Ghz (Python)
194 DPM-VOC+VP 66.25 % 80.45 % 49.86 % 8 s 1 core @ 2.5 Ghz (C/C++)
B. Pepik, M. Stark, P. Gehler and B. Schiele: Multi-view and 3D Deformable Part Models. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2015.
195 BdCost48-25C 65.95 % 78.21 % 51.23 % 4 s 1 core @ 2.5 Ghz (C/C++)
196 65.93 % 78.51 % 57.77 %
197 MDPM-un-BB 64.20 % 77.32 % 50.18 % 60 s 4 core @ 2.5 Ghz (MATLAB)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
198 PDV-Subcat 63.15 % 77.33 % 49.75 % 7 s 1 core @ 2.5 Ghz (C/C++)
J. Shen, X. Zuo, J. Li, W. Yang and H. Ling: A novel pixel neighborhood differential statistic feature for pedestrian and face detection . Pattern Recognition 2017.
199 Lidar_ROI+Yolo(UJS) 62.71 % 70.58 % 55.17 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
200 GNN 62.59 % 76.03 % 50.18 % 0.2 s 1 core @ 2.5 Ghz (Python)
201 MODet
This method makes use of Velodyne laser scans.
 61.84 % 67.21 % 61.57 % 0.05 s GTX1080Ti
202 yl_net 61.01 % 66.08 % 61.29 % 0.03 s GPU @ 2.5 Ghz (Python)
203 DPM-C8B1
This method uses stereo information.
 60.99 % 74.95 % 47.16 % 15 s 4 cores @ 2.5 Ghz (Matlab + C/C++)
J. Yebes, L. Bergasa and M. García-Garrido: Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes. Sensors 2015.
J. Yebes, L. Bergasa, R. Arroyo and A. Lázaro: Supervised learning and evaluation of KITTI's cars detector with DPM. IV 2014.
204 tiny_rfdet code 60.89 % 66.76 % 57.88 % 0.01 s GPU @ 2.5 Ghz (Python)
205 SubCat48LDCF code 60.53 % 78.16 % 43.66 % 0.5 s 8 cores @ 3.5 Ghz (Matlab + C/C++)
A. Fernández-Baldera, J. Buenaposada and L. Baumela: BAdaCost: Multi-class Boosting with Costs . Pattern Recognition 2018.
206 SA_3D
This method makes use of Velodyne laser scans.
 59.14 % 69.84 % 49.62 % 0.3 s 1 core @ 2.5 Ghz (Python)
207 SAMME48LDCF code 58.50 % 76.22 % 47.50 % 0.5 s 8 cores @ 3.5 Ghz (Matlab + C/C++)
A. Fernández-Baldera, J. Buenaposada and L. Baumela: BAdaCost: Multi-class Boosting with Costs . Pattern Recognition 2018.
208 monoref3d 58.27 % 74.15 % 49.71 % 0.1 s 1 core @ 2.5 Ghz (Python)
209 ref3D 58.27 % 74.15 % 49.71 % 0.1 s 1 core @ 2.5 Ghz (Python + C/C++)
210 BirdNet
This method makes use of Velodyne laser scans.
 57.47 % 78.18 % 56.66 % 0.11 s Titan Xp GPU
J. Beltrán, C. Guindel, F. Moreno, D. Cruzado, F. García and A. Escalera: BirdNet: A 3D Object Detection Framework from LiDAR Information. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
211 100Frcnn 57.47 % 81.09 % 48.37 % 2 s 4 cores @ 2.5 Ghz (Python + C/C++)
212 LSVM-MDPM-sv 57.44 % 71.70 % 46.58 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
A. Geiger, C. Wojek and R. Urtasun: Joint 3D Estimation of Objects and Scene Layout. NIPS 2011.
213 ref3D 57.14 % 74.81 % 47.95 % 0.1 s 1 core @ 2.5 Ghz (Python)
214 mylsi-faster-rcnn 56.13 % 79.05 % 48.48 % 0.3 s 1 core @ 2.5 Ghz (Python)
215 LSVM-MDPM-us code 56.10 % 70.52 % 42.87 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
216 ACF-SC 55.76 % 69.76 % 46.27 % <0.3 s 1 core @ >3.5 Ghz (Matlab + C/C++)
C. Cadena, A. Dick and I. Reid: A Fast, Modular Scene Understanding System using Context-Aware Object Detection. Robotics and Automation (ICRA), 2015 IEEE International Conference on 2015.
217 VeloFCN
This method makes use of Velodyne laser scans.
 53.45 % 70.68 % 46.90 % 1 s GPU @ 2.5 Ghz (Python + C/C++)
B. Li, T. Zhang and T. Xia: Vehicle Detection from 3D Lidar Using Fully Convolutional Network. RSS 2016 .
218 ACF 52.81 % 62.82 % 43.89 % 0.2 s 1 core @ >3.5 Ghz (Matlab + C/C++)
P. Doll\'ar, R. Appel, S. Belongie and P. Perona: Fast Feature Pyramids for Object Detection. PAMI 2014.
P. Doll\'ar: Piotr's Image and Video Matlab Toolbox (PMT). .
219 RT3DStereo
This method uses stereo information.
 48.92 % 57.56 % 42.81 % 0.08 s GPU @ 2.5 Ghz (C/C++)
220 TopNet-HighRes
This method makes use of Velodyne laser scans.
 48.87 % 59.77 % 43.15 % 101ms NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
221 DLMB
This method makes use of Velodyne laser scans.
This is an online method (no batch processing).
 48.76 % 59.32 % 43.19 % 0.03 s 8 cores @ 3.5 Ghz (C/C++)
222 Vote3D
This method makes use of Velodyne laser scans.
 48.05 % 56.66 % 42.64 % 0.5 s 4 cores @ 2.8 Ghz (C/C++)
D. Wang and I. Posner: Voting for Voting in Online Point Cloud Object Detection. Proceedings of Robotics: Science and Systems 2015.
223 Multimodal Detection
This method makes use of Velodyne laser scans.
 code 46.77 % 64.04 % 39.38 % 0.06 s GPU @ 3.5 Ghz (Matlab + C/C++)
A. Asvadi, L. Garrote, C. Premebida, P. Peixoto and U. Nunes: Multimodal vehicle detection: fusing 3D- LIDAR and color camera data. Pattern Recognition Letters 2017.
224 softyolo 45.77 % 62.82 % 39.77 % 0.16 s 4 cores @ 2.5 Ghz (Python)
225 rpn 43.99 % 65.47 % 36.33 % 0.01 s 1 core @ 2.5 Ghz (Python)
226 RT3D
This method makes use of Velodyne laser scans.
 39.71 % 49.96 % 41.47 % 0.09 s GPU @ 1.8Ghz
Y. Zeng, Y. Hu, S. Liu, J. Ye, Y. Han, X. Li and N. Sun: RT3D: Real-Time 3-D Vehicle Detection in LiDAR Point Cloud for Autonomous Driving. IEEE Robotics and Automation Letters 2018.
227 KD53-20 37.82 % 52.30 % 32.71 % 0.19 s 4 cores @ 2.5 Ghz (Python)
228 DT3D 35.98 % 49.23 % 31.78 % 0,21s GPU @ 2.5 Ghz (Python)
229 Licar
This method makes use of Velodyne laser scans.
 33.89 % 41.60 % 35.17 % 0.09 s GPU @ 2.0 Ghz (Python)
230 CSoR
This method makes use of Velodyne laser scans.
 code 26.13 % 35.24 % 22.69 % 3.5 s 4 cores @ >3.5 Ghz (Python + C/C++)
L. Plotkin: PyDriver: Entwicklung eines Frameworks für räumliche Detektion und Klassifikation von Objekten in Fahrzeugumgebung. 2015.
231 R-CNN_VGG 26.04 % 32.23 % 20.93 % 10 s GPU @ 2.5 Ghz (Matlab + C/C++)
232 FCN-Depth code 25.66 % 50.55 % 24.95 % 1 s GPU @ 1.5 Ghz (Matlab + C/C++)
233 mBoW
This method makes use of Velodyne laser scans.
 23.76 % 37.63 % 18.44 % 10 s 1 core @ 2.5 Ghz (C/C++)
J. Behley, V. Steinhage and A. Cremers: Laser-based Segment Classification Using a Mixture of Bag-of-Words. Proc. of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2013.
234 DepthCN
This method makes use of Velodyne laser scans.
 code 23.21 % 37.59 % 18.00 % 2.3 s GPU @ 3.5 Ghz (Matlab)
A. Asvadi, L. Garrote, C. Premebida, P. Peixoto and U. Nunes: DepthCN: vehicle detection using 3D- LIDAR and convnet. IEEE ITSC 2017.
235 DLnet 20.30 % 23.46 % 17.96 % 0.3 s 4 cores @ 2.5 Ghz (C/C++)
236 YOLOv2 code 19.31 % 28.37 % 15.94 % 0.02 s GPU @ 3.5 Ghz (C/C++)
J. Redmon, S. Divvala, R. Girshick and A. Farhadi: You only look once: Unified, real-time object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016.
J. Redmon and A. Farhadi: YOLO9000: Better, Faster, Stronger. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2017.
237 TopNet-UncEst
This method makes use of Velodyne laser scans.
 13.77 % 10.35 % 13.49 % 0.09 s NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, M. Braun, M. Lauer and C. Stiller: Capturing Object Detection Uncertainty in Multi-Layer Grid Maps. 2019.
238 TopNet-Retina
This method makes use of Velodyne laser scans.
 6.36 % 7.79 % 6.31 % 52ms GeForce 1080Ti (tensorflow-gpu, v1.12)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
239 TopNet-DecayRate
This method makes use of Velodyne laser scans.
 0.04 % 0.04 % 0.04 % 92 ms NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
240 LaserNet 0.00 % 0.00 % 0.00 % 12 ms GPU @ 2.5 Ghz (C/C++)
G. Meyer, A. Laddha, E. Kee, C. Vallespi-Gonzalez and C. Wellington: LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2019.
241 SN-net 0.00 % 0.00 % 0.00 % 0.8 s GPU @ 2.5 Ghz (Python + C/C++)
242 JSyolo 0.00 % 0.00 % 0.00 % 0.16 s 4 cores @ 2.5 Ghz (Python)
Table as LaTeX | Only published Methods

Pedestrian


Method Setting Code Moderate Easy Hard Runtime Environment
1 Alibaba-CityBrain 80.90 % 88.13 % 74.08 % 1.5 s GPU @ 2.5 Ghz (Python + C/C++)
2 FichaDL 79.68 % 87.76 % 74.93 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
3 ExtAtt 79.63 % 87.95 % 74.78 % 1.2 s GPU @ 2.5 Ghz (Python + C/C++)
4 DH-ARI 78.29 % 87.43 % 69.91 % 3.6 s GPU @ 2.5 Ghz (Python + C/C++)
5 EM-FPS 77.61 % 84.93 % 72.52 % 0.15 s GPU @ 1.5 Ghz (Python + C/C++)
6 F-PointNet
This method makes use of Velodyne laser scans.
 code 77.25 % 87.81 % 74.46 % 0.17 s GPU @ 3.0 Ghz (Python)
C. Qi, W. Liu, C. Wu, H. Su and L. Guibas: Frustum PointNets for 3D Object Detection from RGB-D Data. arXiv preprint arXiv:1711.08488 2017.
7 TuSimple code 77.04 % 86.78 % 72.40 % 1.6 s GPU @ 2.5 Ghz (Python + C/C++)
F. Yang, W. Choi and Y. Lin: Exploit all the layers: Fast and accurate cnn object detector with scale dependent pooling and cascaded rejection classifiers. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016.
K. He, X. Zhang, S. Ren and J. Sun: Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition 2016.
8 THICV-YDM 76.91 % 87.27 % 69.02 % 0.06 s GPU @ 2.5 Ghz (Python)
9 Argus_detection_v1 75.51 % 83.49 % 71.24 % 0.25 s GPU @ 1.5 Ghz (C/C++)
10 RRC code 75.33 % 84.16 % 70.39 % 3.6 s GPU @ 2.5 Ghz (C/C++)
J. Ren, X. Chen, J. Liu, W. Sun, J. Pang, Q. Yan, Y. Tai and L. Xu: Accurate Single Stage Detector Using Recurrent Rolling Convolution. CVPR 2017.
11 VCTNet 75.22 % 85.49 % 71.55 % 0.02 s GPU @ 1.5 Ghz (C/C++)
12 MHN 74.60 % 85.81 % 68.94 % 0.39 s GPU @ 2.5 Ghz (Python)
13 Aston-EAS 74.52 % 85.12 % 69.35 % 0.24 s GPU @ 2.5 Ghz (Python + C/C++)
J. Wei, J. He, Y. Zhou, K. Chen, Z. Tang and Z. Xiong: Enhanced Object Detection With Deep Convolutional Neural Networks for Advanced Driving Assistance. IEEE Transactions on Intelligent Transportation Systems 2019.
14 ECP Faster R-CNN 74.27 % 84.12 % 70.06 % 0.25 s GPU @ 2.5 Ghz (Python)
M. Braun, S. Krebs, F. Flohr and D. Gavrila: The EuroCity Persons Dataset: A Novel Benchmark for Object Detection. CoRR 2018.
15 SJTU-HW 74.24 % 85.42 % 69.34 % 0.85s GPU @ 1.5 Ghz (Python + C/C++)
S. Zhang, X. Zhao, L. Fang, F. Haiping and S. Haitao: LED: LOCALIZATION-QUALITY ESTIMATION EMBEDDED DETECTOR. IEEE International Conference on Image Processing 2018.
L. Fang, X. Zhao and S. Zhang: Small-objectness sensitive detection based on shifted single shot detector. Multimedia Tools and Applications 2018.
16 CLA 74.03 % 84.26 % 68.45 % 0.3 s GPU @ 2.5 Ghz (Matlab + C/C++)
17 Multi-3D
This method makes use of Velodyne laser scans.
 74.00 % 84.01 % 68.74 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
18 ECV-NET 73.74 % 84.58 % 66.35 % 0.4 s GPU @ 2.5 Ghz (C/C++)
19 BOE_IOT_AIBD 73.73 % 84.67 % 68.71 % 0.8 s GPU @ 2.5 Ghz (Python)
20 MS-CNN code 73.62 % 83.70 % 68.28 % 0.4 s GPU @ 2.5 Ghz (C/C++)
Z. Cai, Q. Fan, R. Feris and N. Vasconcelos: A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection. ECCV 2016.
21 SAITv1 72.61 % 84.79 % 67.94 % 0.15 s GPU @ 2.5 Ghz (C/C++)
22 F-ConvNet
This method makes use of Velodyne laser scans.
 72.37 % 79.98 % 66.61 % 0.47 s GPU @ 2.5 Ghz (Python + C/C++)
Z. Wang and K. Jia: Frustum ConvNet: Sliding Frustums to Aggregate Local Point-Wise Features for Amodal 3D Object Detection. arXiv preprint arXiv:1903.01864 2019.
23 Sogo_MM 71.84 % 83.45 % 67.00 % 1.5 s GPU @ 2.5 Ghz (C/C++)
24 GN 71.55 % 80.73 % 64.82 % 1 s GPU @ 2.5 Ghz (Matlab + C/C++)
S. Jung and K. Hong: Deep network aided by guiding network for pedestrian detection. Pattern Recognition Letters 2017.
25 SubCNN 71.34 % 83.17 % 66.36 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
Y. Xiang, W. Choi, Y. Lin and S. Savarese: Subcategory-aware Convolutional Neural Networks for Object Proposals and Detection. IEEE Winter Conference on Applications of Computer Vision (WACV) 2017.
26 VMVS
This method makes use of Velodyne laser scans.
 70.89 % 81.11 % 67.23 % 0.25 s GPU @ 2.5 Ghz (Python)
27 IVA code 70.63 % 83.03 % 64.68 % 0.4 s GPU @ 2.5 Ghz (C/C++)
Y. Zhu, J. Wang, C. Zhao, H. Guo and H. Lu: Scale-adaptive Deconvolutional Regression Network for Pedestrian Detection. ACCV 2016.
S. Ren, K. He, R. Girshick and J. Sun: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 2015.
28 SDP+RPN 70.20 % 79.98 % 64.84 % 0.4 s GPU @ 2.5 Ghz (Python + C/C++)
F. Yang, W. Choi and Y. Lin: Exploit All the Layers: Fast and Accurate CNN Object Detector with Scale Dependent Pooling and Cascaded Rejection Classifiers. Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition 2016.
S. Ren, K. He, R. Girshick and J. Sun: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 2015.
29 MM-MRFC
This method uses optical flow information.
This method makes use of Velodyne laser scans.
 69.96 % 82.37 % 64.76 % 0.05 s GPU @ 2.5 Ghz (C/C++)
A. Costea, R. Varga and S. Nedevschi: Fast Boosting based Detection using Scale Invariant Multimodal Multiresolution Filtered Features. CVPR 2017.
30 TridentNet 69.58 % 81.27 % 64.18 % 0.2 s GPU @ 2.5 Ghz (Python)
31 MDC
This method makes use of Velodyne laser scans.
 69.58 % 86.37 % 68.44 % 0.17 s GPU @ 2.5 Ghz (Python)
32 MonoPSR 68.91 % 85.93 % 60.83 % 0.2 s GPU @ 3.5 Ghz (Python)
J. Ku*, A. Pon* and S. Waslander: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. CVPR 2019.
33 YOLOv3.5 68.33 % 79.61 % 60.85 % 0.05 s GPU @ 2.5 Ghz (Python)
34 HBA-RCNN 68.26 % 77.76 % 62.86 % 0.4 s 1 core @ 2.5 Ghz (C/C++)
35 DA 67.89 % 79.91 % 64.83 % 0.08 s 1 core @ 2.5 Ghz (C/C++)
36 3DOP
This method uses stereo information.
 code 67.46 % 82.36 % 64.71 % 3s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Y. Zhu, A. Berneshawi, H. Ma, S. Fidler and R. Urtasun: 3D Object Proposals for Accurate Object Class Detection. NIPS 2015.
37 DeepStereoOP 67.32 % 82.50 % 65.14 % 3.4 s GPU @ 3.5 Ghz (Matlab + C/C++)
C. Pham and J. Jeon: Robust Object Proposals Re-ranking for Object Detection in Autonomous Driving Using Convolutional Neural Networks. Signal Processing: Image Communiation 2017.
38 sensekitti code 67.28 % 80.12 % 62.25 % 4.5 s GPU @ 2.5 Ghz (Python + C/C++)
B. Yang, J. Yan, Z. Lei and S. Li: Craft Objects from Images. CVPR 2016.
39 ODES code 67.25 % 77.95 % 62.28 % 0.02 s GPU @ 2.5 Ghz (Python)
ERROR: Wrong syntax in BIBTEX file.
40 Mono3D code 66.66 % 77.30 % 63.44 % 4.2 s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Z. Zhang, H. Ma, S. Fidler and R. Urtasun: Monocular 3D Object Detection for Autonomous Driving. CVPR 2016.
41 Faster R-CNN code 65.91 % 78.35 % 61.19 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
S. Ren, K. He, R. Girshick and J. Sun: Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks. NIPS 2015.
42 AtrousDet 65.18 % 77.19 % 58.14 % 0.05 s TITAN X
43 SDP+CRC (ft) 64.25 % 77.81 % 59.31 % 0.6 s GPU @ 2.5 Ghz (C/C++)
F. Yang, W. Choi and Y. Lin: Exploit All the Layers: Fast and Accurate CNN Object Detector with Scale Dependent Pooling and Cascaded Rejection Classifiers. Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition 2016.
44 PCN 63.48 % 77.88 % 58.59 % 0.6 s
45 Pose-RCNN 63.38 % 77.69 % 57.42 % 2 s >8 cores @ 2.5 Ghz (Python)
M. Braun, Q. Rao, Y. Wang and F. Flohr: Pose-RCNN: Joint object detection and pose estimation using 3D object proposals. Intelligent Transportation Systems (ITSC), 2016 IEEE 19th International Conference on 2016.
46 CFM 63.26 % 74.21 % 56.44 % <2 s GPU @ 2.5 Ghz (Matlab + C/C++)
Q. Hu, P. Wang, C. Shen, A. Hengel and F. Porikli: Pushing the Limits of Deep CNNs for Pedestrian Detection. IEEE Transactions on Circuits and Systems for Video Technology 2017.
47 IPOD 63.07 % 73.28 % 56.71 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
48 ALV303 61.77 % 69.13 % 54.54 % 0.2 s GPU @ 2.5 Ghz (Matlab + C/C++)
49 RPN+BF code 61.29 % 75.58 % 56.08 % 0.6 s GPU @ 2.5 Ghz (Matlab + C/C++)
L. Zhang, L. Lin, X. Liang and K. He: Is Faster R-CNN Doing Well for Pedestrian Detection?. ECCV 2016.
50 ReSqueeze 61.25 % 72.78 % 57.43 % 0.03 s GPU @ >3.5 Ghz (Python)
51 Regionlets 61.16 % 72.96 % 55.22 % 1 s >8 cores @ 2.5 Ghz (C/C++)
X. Wang, M. Yang, S. Zhu and Y. Lin: Regionlets for Generic Object Detection. T-PAMI 2015.
W. Zou, X. Wang, M. Sun and Y. Lin: Generic Object Detection with Dense Neural Patterns and Regionlets. British Machine Vision Conference 2014.
C. Long, X. Wang, G. Hua, M. Yang and Y. Lin: Accurate Object Detection with Location Relaxation and Regionlets Relocalization. Asian Conference on Computer Vision 2014.
52 merge12-12 60.66 % 78.15 % 58.67 % 0.2 s 4 cores @ 2.5 Ghz (Python)
53 cascadercnn 60.64 % 77.88 % 52.69 % 0.36 s 4 cores @ 2.5 Ghz (Python)
54 cas+res+soft 60.60 % 77.96 % 58.56 % 0.2 s 4 cores @ 2.5 Ghz (Python)
55 bin 60.54 % 70.13 % 56.55 % 15ms s GPU @ >3.5 Ghz (Python)
56 cas_retina 60.30 % 77.71 % 58.34 % 0.2 s 4 cores @ 2.5 Ghz (Python)
57 A-VoxelNet 59.98 % 69.26 % 58.48 % 0.029 s GPU @ 2.5 Ghz (Python)
58 cas_retina_1_13 59.87 % 77.11 % 57.81 % 0.03 s 4 cores @ 2.5 Ghz (Python)
59 anm 59.21 % 75.51 % 56.49 % 3 s 1 core @ 2.5 Ghz (C/C++)
60 CompACT-Deep 58.73 % 69.70 % 52.69 % 1 s 1 core @ 2.5 Ghz (Matlab + C/C++)
Z. Cai, M. Saberian and N. Vasconcelos: Learning Complexity-Aware Cascades for Deep Pedestrian Detection. ICCV 2015.
61 DeepParts 58.68 % 70.46 % 52.73 % ~1 s GPU @ 2.5 Ghz (Matlab)
Y. Tian, P. Luo, X. Wang and X. Tang: Deep Learning Strong Parts for Pedestrian Detection. ICCV 2015.
62 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 58.42 % 67.32 % 57.44 % 0.1 s Titan X (Pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
63 LPN 58.18 % 70.54 % 54.18 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
64 SA_3D 57.85 % 68.55 % 50.45 % 0.3 s GPU @ 2.5 Ghz (Python)
65 mymask-rcnn 57.79 % 73.18 % 55.11 % 0.3 s 1 core @ 2.5 Ghz (Python)
66 SA_3D
This method makes use of Velodyne laser scans.
 57.40 % 68.11 % 50.09 % 0.3 s 1 core @ 2.5 Ghz (Python)
67 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 57.23 % 66.08 % 55.10 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
68 FilteredICF 57.12 % 69.05 % 51.46 % ~ 2 s >8 cores @ 2.5 Ghz (Matlab + C/C++)
S. Zhang, R. Benenson and B. Schiele: Filtered Channel Features for Pedestrian Detection. CVPR 2015.
69 ZKNet 57.11 % 70.20 % 52.15 % 0.01 s GPU @ 2.0 Ghz (Python)
70 RFCN 56.91 % 71.17 % 50.06 % 0.2 s 4 cores @ 2.5 Ghz (Python)
71 FRCNN+Or code 56.78 % 71.18 % 52.86 % 0.09 s Titan Xp GPU
C. Guindel, D. Martin and J. Armingol: Fast Joint Object Detection and Viewpoint Estimation for Traffic Scene Understanding. IEEE Intelligent Transportation Systems Magazine 2018.
C. Guindel, D. Martin and J. Armingol: Joint Object Detection and Viewpoint Estimation using CNN features. IEEE International Conference on Vehicular Electronics and Safety (ICVES) 2017.
72 FD2 56.68 % 71.09 % 51.65 % 0.01 s GPU @ >3.5 Ghz (Python + C/C++)
73 MV-RGBD-RF
This method makes use of Velodyne laser scans.
 56.59 % 73.05 % 49.63 % 4 s 4 cores @ 2.5 Ghz (C/C++)
A. Gonzalez, D. Vazquez, A. Lopez and J. Amores: On-Board Object Detection: Multicue, Multimodal, and Multiview Random Forest of Local Experts.. IEEE Trans. on Cybernetics 2016.
A. Gonzalez, G. Villalonga, J. Xu, D. Vazquez, J. Amores and A. Lopez: Multiview Random Forest of Local Experts Combining RGB and LIDAR data for Pedestrian Detection. IEEE Intelligent Vehicles Symposium (IV) 2015.
74 CHTTL MMF 56.01 % 73.22 % 50.26 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
75 RFCN_RFB 55.86 % 69.32 % 49.18 % 0.2 s 4 cores @ 2.5 Ghz (Python)
76 SECOND code 55.74 % 65.73 % 49.08 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
77 PointPillars
This method makes use of Velodyne laser scans.
 code 55.68 % 64.66 % 53.93 % 16 ms 1080ti GPU and Intel i7 CPU
A. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang and O. Beijbom: PointPillars: Fast Encoders for Object Detection from Point Clouds. CVPR 2019.
78 yolo800 55.49 % 71.11 % 53.92 % 0.13 s 4 cores @ 2.5 Ghz (Python)
79 STD 55.44 % 69.09 % 53.46 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
80 Vote3Deep
This method makes use of Velodyne laser scans.
 55.38 % 67.94 % 52.62 % 1.5 s 4 cores @ 2.5 Ghz (C/C++)
M. Engelcke, D. Rao, D. Zeng Wang, C. Hay Tong and I. Posner: Vote3Deep: Fast Object Detection in 3D Point Clouds Using Efficient Convolutional Neural Networks. ArXiv e-prints 2016.
81 CONV-BOX
This method makes use of Velodyne laser scans.
 55.23 % 63.98 % 54.18 % 0.2 s Tesla V100
82 epBRM
This method makes use of Velodyne laser scans.
 54.62 % 62.46 % 53.30 % 0.10 s 1 core @ 2.5 Ghz (C/C++)
83 TAFT 54.59 % 67.07 % 48.48 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
J. Shen, X. Zuo, W. Yang, D. Prokhorov, X. Mei and H. Ling: Differential Features for Pedestrian Detection: A Taylor Series Perspective. IEEE Transactions on Intelligent Transportation Systems 2018.
84 pAUCEnsT 54.58 % 66.11 % 48.49 % 60 s 1 core @ 2.5 Ghz (Matlab + C/C++)
S. Paisitkriangkrai, C. Shen and A. Hengel: Pedestrian Detection with Spatially Pooled Features and Structured Ensemble Learning. arXiv 2014.
85 SCANet 54.02 % 64.57 % 48.05 % 0.17 s >8 cores @ 2.5 Ghz (Python)
86 NM code 53.98 % 69.06 % 50.76 % 0.01 s GPU @ 2.5 Ghz (Python)
87 fasterrcnn 53.80 % 69.00 % 51.35 % 0.2 s 4 cores @ 2.5 Ghz (Python)
88 NEUAV 53.75 % 68.86 % 48.04 % 0.06 s GPU @ 2.5 Ghz (Python)
89 PDV2 53.74 % 65.71 % 49.47 % 3.7 s 1 core @ 3.0 Ghz Matlab (C/C++)
J. Shen, X. Zuo, J. Li, W. Yang and H. Ling: A novel pixel neighborhood differential statistic feature for pedestrian and face detection . Pattern Recognition 2017.
90 PP_v1.0 code 53.59 % 62.16 % 51.51 % 0.02s 1 core @ 2.5 Ghz (C/C++)
91 Shift R-CNN 53.33 % 71.11 % 44.71 % 0.25 s GPU @ 1.5 Ghz (Python)
A. Naiden, V. Paunescu, G. Kim, B. Jeon and M. Leordeanu: Shift R-CNN: Deep Monocular 3D Object Detection With Closed-form Geometric Constraints. ICIP 2019.
92 MTDP 52.97 % 66.97 % 47.64 % 0.15 s GPU @ 2.0 Ghz (Python)
93 detectron code 52.42 % 69.89 % 51.70 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
94 cascade_gw 51.10 % 67.58 % 43.34 % 0.2 s 4 cores @ 2.5 Ghz (Python)
95 YOLOv3+d 51.03 % 67.23 % 48.87 % 0.04 s GPU @ 1.5 Ghz (C/C++)
96 ACFD
This method makes use of Velodyne laser scans.
 code 50.91 % 61.59 % 45.51 % 0.2 s 4 cores @ >3.5 Ghz (C/C++)
M. Dimitrievski, P. Veelaert and W. Philips: Semantically aware multilateral filter for depth upsampling in automotive LiDAR point clouds. IEEE Intelligent Vehicles Symposium, IV 2017, Los Angeles, CA, USA, June 11-14, 2017 2017.
97 PointRCNN
This method makes use of Velodyne laser scans.
 code 50.88 % 59.05 % 48.46 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
S. Shi, X. Wang and H. Li: PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud. CVPR 2019.
98 myfaster-rcnn-v1.5 50.62 % 66.91 % 48.16 % 0.1 s 1 core @ 2.5 Ghz (Python)
99 CLF3D
This method makes use of Velodyne laser scans.
 50.25 % 66.10 % 48.66 % 0.13 s GPU @ 2.5 Ghz (Python)
100 R-CNN 50.20 % 62.05 % 44.85 % 4 s GPU @ 3.3 Ghz (C/C++)
J. Hosang, M. Omran, R. Benenson and B. Schiele: Taking a Deeper Look at Pedestrians. arXiv 2015.
101 SS3D 49.81 % 59.46 % 42.44 % 48 ms Tesla V100 (Python)
102 SeRC 49.81 % 65.31 % 42.40 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
103 Resnet101Faster rcnn 49.64 % 64.97 % 48.47 % 1 s 1 core @ 2.5 Ghz (Python)
104 mylsi-faster-rcnn 49.31 % 64.64 % 45.96 % 0.3 s 1 core @ 2.5 Ghz (Python)
105 Int-YOLO code 48.93 % 64.40 % 48.02 % 0.03 s 1 core @ 2.5 Ghz (C/C++)
ERROR: Wrong syntax in BIBTEX file.
106 ARPNET 48.77 % 63.16 % 47.65 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
107 ELLIOT
This method makes use of Velodyne laser scans.
 48.26 % 59.05 % 45.52 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
108 CFR
This method makes use of Velodyne laser scans.
 48.16 % 63.07 % 47.51 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
109 ACF 47.29 % 60.11 % 42.90 % 1 s 1 core @ 3.5 Ghz (Matlab + C/C++)
P. Doll\'ar, R. Appel, S. Belongie and P. Perona: Fast Feature Pyramids for Object Detection. PAMI 2014.
110 Fusion-DPM
This method makes use of Velodyne laser scans.
 code 46.67 % 59.38 % 42.05 % ~ 30 s 1 core @ 3.5 Ghz (Matlab + C/C++)
C. Premebida, J. Carreira, J. Batista and U. Nunes: Pedestrian Detection Combining RGB and Dense LIDAR Data. IROS 2014.
111 ACF-MR 46.23 % 58.85 % 42.10 % 0.6 s 1 core @ 3.5 Ghz (C/C++)
R. Rajaram, E. Ohn-Bar and M. Trivedi: Looking at Pedestrians at Different Scales: A Multi-resolution Approach and Evaluations. T-ITS 2016.
112 HA-SSVM 45.51 % 58.91 % 41.08 % 21 s 1 core @ >3.5 Ghz (Matlab + C/C++)
J. Xu, S. Ramos, D. Vázquez and A. López: Hierarchical Adaptive Structural SVM for Domain Adaptation. IJCV 2016.
113 DPM-VOC+VP 44.86 % 59.60 % 40.37 % 8 s 1 core @ 2.5 Ghz (C/C++)
B. Pepik, M. Stark, P. Gehler and B. Schiele: Multi-view and 3D Deformable Part Models. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2015.
114 Cmerge 44.81 % 62.62 % 44.53 % 0.2 s 4 cores @ 2.5 Ghz (Python)
115 ACF-SC 44.77 % 54.20 % 39.57 % <0.3 s 1 core @ >3.5 Ghz (Matlab + C/C++)
C. Cadena, A. Dick and I. Reid: A Fast, Modular Scene Understanding System using Context-Aware Object Detection. Robotics and Automation (ICRA), 2015 IEEE International Conference on 2015.
116 SquaresICF code 44.42 % 57.47 % 40.08 % 1 s GPU @ >3.5 Ghz (C/C++)
R. Benenson, M. Mathias, T. Tuytelaars and L. Gool: Seeking the strongest rigid detector. CVPR 2013.
117 AVOD
This method makes use of Velodyne laser scans.
 code 43.49 % 51.64 % 37.79 % 0.08 s Titan X (pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
118 Retinanet100 42.83 % 52.43 % 35.02 % 0.2 s 4 cores @ 2.5 Ghz (Python)
119 GNN 42.56 % 58.22 % 40.53 % 0.2 s 1 core @ 2.5 Ghz (Python)
120 SubCat 42.34 % 54.06 % 37.95 % 1.2 s 6 cores @ 2.5 Ghz (Matlab + C/C++)
E. Ohn-Bar and M. Trivedi: Fast and Robust Object Detection Using Visual Subcategories. Computer Vision and Pattern Recognition Workshops Mobile Vision 2014.
121 myfaster-rcnn 41.91 % 57.22 % 39.62 % 0.01 s 1 core @ 2.5 Ghz (Python)
122 yolov3_warp 41.07 % 56.07 % 39.08 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
ERROR: Wrong syntax in BIBTEX file.
123 softyolo 40.78 % 55.95 % 39.57 % 0.16 s 4 cores @ 2.5 Ghz (Python)
124 ACF 40.62 % 49.08 % 36.66 % 0.2 s 1 core @ >3.5 Ghz (Matlab + C/C++)
P. Doll\'ar, R. Appel, S. Belongie and P. Perona: Fast Feature Pyramids for Object Detection. PAMI 2014.
P. Doll\'ar: Piotr's Image and Video Matlab Toolbox (PMT). .
125 multi-task CNN 40.34 % 51.38 % 34.98 % 25.1 ms GPU @ 2.0 Ghz (Python)
M. Oeljeklaus, F. Hoffmann and T. Bertram: A Fast Multi-Task CNN for Spatial Understanding of Traffic Scenes. IEEE Intelligent Transportation Systems Conference 2018.
126 KD53-20 39.90 % 47.15 % 35.32 % 0.19 s 4 cores @ 2.5 Ghz (Python)
127 LSVM-MDPM-sv 39.36 % 51.75 % 35.95 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
A. Geiger, C. Wojek and R. Urtasun: Joint 3D Estimation of Objects and Scene Layout. NIPS 2011.
128 pedestrian_cnn 39.07 % 53.60 % 37.91 % 1 s 1 core @ 2.5 Ghz (C/C++)
129 Lidar_ROI+Yolo(UJS) 38.76 % 47.11 % 32.33 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
130 LSVM-MDPM-us code 38.35 % 50.01 % 34.78 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
131 37.45 % 45.89 % 35.08 %
132 X_MD 37.38 % 50.17 % 36.49 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
133 anonymous
This method makes use of Velodyne laser scans.
 36.65 % 49.15 % 36.18 % 0.75 s GPU @ 3.5 Ghz (C/C++)
134 Complexer-YOLO
This method makes use of Velodyne laser scans.
 36.10 % 42.63 % 35.57 % 0.06 s GPU @ 3.5 Ghz (C/C++)
135 Vote3D
This method makes use of Velodyne laser scans.
 35.74 % 44.47 % 33.72 % 0.5 s 4 cores @ 2.8 Ghz (C/C++)
D. Wang and I. Posner: Voting for Voting in Online Point Cloud Object Detection. Proceedings of Robotics: Science and Systems 2015.
136 rpn 32.79 % 46.95 % 31.70 % 0.01 s 1 core @ 2.5 Ghz (Python)
137 RT3DStereo
This method uses stereo information.
 32.01 % 44.54 % 31.50 % 0.08 s GPU @ 2.5 Ghz (C/C++)
138 mBoW
This method makes use of Velodyne laser scans.
 31.37 % 44.36 % 30.62 % 10 s 1 core @ 2.5 Ghz (C/C++)
J. Behley, V. Steinhage and A. Cremers: Laser-based Segment Classification Using a Mixture of Bag-of-Words. Proc. of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2013.
139 BirdNet
This method makes use of Velodyne laser scans.
 30.90 % 36.83 % 29.93 % 0.11 s Titan Xp GPU
J. Beltrán, C. Guindel, F. Moreno, D. Cruzado, F. García and A. Escalera: BirdNet: A 3D Object Detection Framework from LiDAR Information. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
140 DPM-C8B1
This method uses stereo information.
 29.03 % 38.96 % 25.61 % 15 s 4 cores @ 2.5 Ghz (Matlab + C/C++)
J. Yebes, L. Bergasa and M. García-Garrido: Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes. Sensors 2015.
J. Yebes, L. Bergasa, R. Arroyo and A. Lázaro: Supervised learning and evaluation of KITTI's cars detector with DPM. IV 2014.
141 100Frcnn 26.73 % 35.65 % 26.46 % 2 s 4 cores @ 2.5 Ghz (Python + C/C++)
142 R-CNN_VGG 23.16 % 28.95 % 22.17 % 10 s GPU @ 2.5 Ghz (Matlab + C/C++)
143 TopNet-Retina
This method makes use of Velodyne laser scans.
 19.67 % 25.17 % 16.33 % 52ms GeForce 1080Ti (tensorflow-gpu, v1.12)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
144 DT3D 19.19 % 27.02 % 18.98 % 0,21s GPU @ 2.5 Ghz (Python)
145 TopNet-HighRes
This method makes use of Velodyne laser scans.
 17.57 % 22.98 % 17.35 % 101ms NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
146 YOLOv2 code 16.19 % 20.80 % 15.43 % 0.02 s GPU @ 3.5 Ghz (C/C++)
J. Redmon, S. Divvala, R. Girshick and A. Farhadi: You only look once: Unified, real-time object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016.
J. Redmon and A. Farhadi: YOLO9000: Better, Faster, Stronger. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2017.
147 BIP-HETERO 13.38 % 14.85 % 13.25 % ~2 s 1 core @ 2.5 Ghz (C/C++)
A. Mekonnen, F. Lerasle, A. Herbulot and C. Briand: People Detection with Heterogeneous Features and Explicit Optimization on Computation Time. Pattern Recognition (ICPR), 2014 22nd International Conference on 2014.
148 TopNet-UncEst
This method makes use of Velodyne laser scans.
 10.91 % 15.55 % 10.05 % 0.09 s NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, M. Braun, M. Lauer and C. Stiller: Capturing Object Detection Uncertainty in Multi-Layer Grid Maps. 2019.
149 softretina 0.93 % 0.68 % 0.95 % 0.16 s 4 cores @ 2.5 Ghz (Python)
150 JSyolo 0.44 % 0.35 % 0.45 % 0.16 s 4 cores @ 2.5 Ghz (Python)
151 TopNet-DecayRate
This method makes use of Velodyne laser scans.
 0.04 % 0.02 % 0.05 % 92 ms NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
152 SN-net 0.00 % 0.00 % 0.00 % 0.8 s GPU @ 2.5 Ghz (Python + C/C++)
Table as LaTeX | Only published Methods

Cyclist


Method Setting Code Moderate Easy Hard Runtime Environment
1 EM-FPS 79.24 % 84.28 % 71.22 % 0.15 s GPU @ 1.5 Ghz (Python + C/C++)
2 RRC code 76.49 % 84.96 % 65.46 % 3.6 s GPU @ 2.5 Ghz (C/C++)
J. Ren, X. Chen, J. Liu, W. Sun, J. Pang, Q. Yan, Y. Tai and L. Xu: Accurate Single Stage Detector Using Recurrent Rolling Convolution. CVPR 2017.
3 F-ConvNet
This method makes use of Velodyne laser scans.
 76.18 % 84.75 % 67.55 % 0.47 s GPU @ 2.5 Ghz (Python + C/C++)
Z. Wang and K. Jia: Frustum ConvNet: Sliding Frustums to Aggregate Local Point-Wise Features for Amodal 3D Object Detection. arXiv preprint arXiv:1903.01864 2019.
4 VCTNet 75.91 % 83.20 % 67.81 % 0.02 s GPU @ 1.5 Ghz (C/C++)
5 SAITv1 75.83 % 83.99 % 66.45 % 0.15 s GPU @ 2.5 Ghz (C/C++)
6 FichaDL 75.32 % 84.49 % 66.27 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
7 Multi-3D
This method makes use of Velodyne laser scans.
 74.88 % 82.13 % 65.53 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
8 CLA 74.68 % 82.42 % 65.11 % 0.3 s GPU @ 2.5 Ghz (Matlab + C/C++)
9 MS-CNN code 74.45 % 82.34 % 64.91 % 0.4 s GPU @ 2.5 Ghz (C/C++)
Z. Cai, Q. Fan, R. Feris and N. Vasconcelos: A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection. ECCV 2016.
10 TuSimple code 74.26 % 81.38 % 64.88 % 1.6 s GPU @ 2.5 Ghz (Python + C/C++)
F. Yang, W. Choi and Y. Lin: Exploit all the layers: Fast and accurate cnn object detector with scale dependent pooling and cascaded rejection classifiers. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016.
K. He, X. Zhang, S. Ren and J. Sun: Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition 2016.
11 ExtAtt 74.25 % 84.04 % 65.03 % 1.2 s GPU @ 2.5 Ghz (Python + C/C++)
12 Deep3DBox 73.48 % 82.65 % 64.11 % 1.5 s GPU @ 2.5 Ghz (C/C++)
A. Mousavian, D. Anguelov, J. Flynn and J. Kosecka: 3D Bounding Box Estimation Using Deep Learning and Geometry. CVPR 2017.
13 SDP+RPN 73.08 % 81.05 % 64.88 % 0.4 s GPU @ 2.5 Ghz (Python + C/C++)
F. Yang, W. Choi and Y. Lin: Exploit All the Layers: Fast and Accurate CNN Object Detector with Scale Dependent Pooling and Cascaded Rejection Classifiers. Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition 2016.
S. Ren, K. He, R. Girshick and J. Sun: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 2015.
14 PointRCNN
This method makes use of Velodyne laser scans.
 code 72.94 % 83.64 % 66.07 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
S. Shi, X. Wang and H. Li: PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud. CVPR 2019.
15 ECV-NET 72.73 % 82.62 % 62.82 % 0.4 s GPU @ 2.5 Ghz (C/C++)
16 STD 72.63 % 82.18 % 65.16 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
17 sensekitti code 72.50 % 81.76 % 64.00 % 4.5 s GPU @ 2.5 Ghz (Python + C/C++)
B. Yang, J. Yan, Z. Lei and S. Li: Craft Objects from Images. CVPR 2016.
18 F-PointNet
This method makes use of Velodyne laser scans.
 code 72.25 % 84.90 % 65.14 % 0.17 s GPU @ 3.0 Ghz (Python)
C. Qi, W. Liu, C. Wu, H. Su and L. Guibas: Frustum PointNets for 3D Object Detection from RGB-D Data. arXiv preprint arXiv:1711.08488 2017.
19 BOE_IOT_AIBD 71.61 % 82.63 % 63.67 % 0.8 s GPU @ 2.5 Ghz (Python)
20 SubCNN 70.77 % 77.82 % 62.71 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
Y. Xiang, W. Choi, Y. Lin and S. Savarese: Subcategory-aware Convolutional Neural Networks for Object Proposals and Detection. IEEE Winter Conference on Applications of Computer Vision (WACV) 2017.
21 Sogo_MM 70.72 % 77.57 % 62.23 % 1.5 s GPU @ 2.5 Ghz (C/C++)
22 ARPNET 69.94 % 79.83 % 63.08 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
23 ODES code 69.80 % 78.51 % 61.32 % 0.02 s GPU @ 2.5 Ghz (Python)
ERROR: Wrong syntax in BIBTEX file.
24 TridentNet 69.06 % 80.64 % 60.06 % 0.2 s GPU @ 2.5 Ghz (Python)
25 MonoPSR 68.99 % 79.80 % 60.19 % 0.2 s GPU @ 3.5 Ghz (Python)
J. Ku*, A. Pon* and S. Waslander: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. CVPR 2019.
26 MDC
This method makes use of Velodyne laser scans.
 68.84 % 79.81 % 60.24 % 0.17 s GPU @ 2.5 Ghz (Python)
27 3DOP
This method uses stereo information.
 code 68.81 % 80.17 % 61.36 % 3s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Y. Zhu, A. Berneshawi, H. Ma, S. Fidler and R. Urtasun: 3D Object Proposals for Accurate Object Class Detection. NIPS 2015.
28 PointPillars
This method makes use of Velodyne laser scans.
 code 68.57 % 82.59 % 62.37 % 16 ms 1080ti GPU and Intel i7 CPU
A. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang and O. Beijbom: PointPillars: Fast Encoders for Object Detection from Point Clouds. CVPR 2019.
29 Pose-RCNN 68.04 % 80.19 % 59.95 % 2 s >8 cores @ 2.5 Ghz (Python)
M. Braun, Q. Rao, Y. Wang and F. Flohr: Pose-RCNN: Joint object detection and pose estimation using 3D object proposals. Intelligent Transportation Systems (ITSC), 2016 IEEE 19th International Conference on 2016.
30 Vote3Deep
This method makes use of Velodyne laser scans.
 67.96 % 76.49 % 62.88 % 1.5 s 4 cores @ 2.5 Ghz (C/C++)
M. Engelcke, D. Rao, D. Zeng Wang, C. Hay Tong and I. Posner: Vote3Deep: Fast Object Detection in 3D Point Clouds Using Efficient Convolutional Neural Networks. ArXiv e-prints 2016.
31 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 67.37 % 81.56 % 61.28 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
32 IVA code 67.36 % 77.63 % 59.62 % 0.4 s GPU @ 2.5 Ghz (C/C++)
Y. Zhu, J. Wang, C. Zhao, H. Guo and H. Lu: Scale-adaptive Deconvolutional Regression Network for Pedestrian Detection. ACCV 2016.
S. Ren, K. He, R. Girshick and J. Sun: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 2015.
33 A-VoxelNet 67.13 % 80.77 % 60.37 % 0.029 s GPU @ 2.5 Ghz (Python)
34 epBRM
This method makes use of Velodyne laser scans.
 66.04 % 77.78 % 60.39 % 0.10 s 1 core @ 2.5 Ghz (C/C++)
35 DeepStereoOP 65.72 % 77.00 % 57.74 % 3.4 s GPU @ 3.5 Ghz (Matlab + C/C++)
C. Pham and J. Jeon: Robust Object Proposals Re-ranking for Object Detection in Autonomous Driving Using Convolutional Neural Networks. Signal Processing: Image Communiation 2017.
36 IPOD 65.28 % 82.90 % 57.63 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
37 Mono3D code 63.85 % 75.22 % 58.96 % 4.2 s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Z. Zhang, H. Ma, S. Fidler and R. Urtasun: Monocular 3D Object Detection for Autonomous Driving. CVPR 2016.
38 CONV-BOX
This method makes use of Velodyne laser scans.
 63.84 % 72.62 % 56.69 % 0.2 s Tesla V100
39 DA 63.58 % 79.36 % 56.80 % 0.08 s 1 core @ 2.5 Ghz (C/C++)
40 SCANet 63.26 % 73.72 % 56.41 % 0.17 s >8 cores @ 2.5 Ghz (Python)
41 AtrousDet 62.85 % 76.07 % 55.12 % 0.05 s TITAN X
42 Faster R-CNN code 62.81 % 71.41 % 55.44 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
S. Ren, K. He, R. Girshick and J. Sun: Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks. NIPS 2015.
43 cas+res+soft 60.88 % 75.24 % 53.58 % 0.2 s 4 cores @ 2.5 Ghz (Python)
44 SDP+CRC (ft) 60.87 % 74.31 % 53.95 % 0.6 s GPU @ 2.5 Ghz (C/C++)
F. Yang, W. Choi and Y. Lin: Exploit All the Layers: Fast and Accurate CNN Object Detector with Scale Dependent Pooling and Cascaded Rejection Classifiers. Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition 2016.
45 merge12-12 60.83 % 75.12 % 53.69 % 0.2 s 4 cores @ 2.5 Ghz (Python)
46 ELLIOT
This method makes use of Velodyne laser scans.
 60.04 % 77.40 % 55.15 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
47 PP_v1.0 code 59.92 % 75.52 % 53.73 % 0.02s 1 core @ 2.5 Ghz (C/C++)
48 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 59.32 % 68.65 % 55.82 % 0.1 s Titan X (Pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
49 SECOND code 58.94 % 81.96 % 57.20 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
50 Regionlets 58.69 % 70.09 % 51.81 % 1 s >8 cores @ 2.5 Ghz (C/C++)
X. Wang, M. Yang, S. Zhu and Y. Lin: Regionlets for Generic Object Detection. T-PAMI 2015.
W. Zou, X. Wang, M. Sun and Y. Lin: Generic Object Detection with Dense Neural Patterns and Regionlets. British Machine Vision Conference 2014.
C. Long, X. Wang, G. Hua, M. Yang and Y. Lin: Accurate Object Detection with Location Relaxation and Regionlets Relocalization. Asian Conference on Computer Vision 2014.
51 YOLOv3.5 58.23 % 77.33 % 50.68 % 0.05 s GPU @ 2.5 Ghz (Python)
52 CFR
This method makes use of Velodyne laser scans.
 58.19 % 74.83 % 56.15 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
53 cascadercnn 58.09 % 75.56 % 50.19 % 0.36 s 4 cores @ 2.5 Ghz (Python)
54 Complexer-YOLO
This method makes use of Velodyne laser scans.
 57.53 % 65.82 % 57.47 % 0.06 s GPU @ 3.5 Ghz (C/C++)
55 FRCNN+Or code 57.37 % 70.05 % 51.00 % 0.09 s Titan Xp GPU
C. Guindel, D. Martin and J. Armingol: Fast Joint Object Detection and Viewpoint Estimation for Traffic Scene Understanding. IEEE Intelligent Transportation Systems Magazine 2018.
C. Guindel, D. Martin and J. Armingol: Joint Object Detection and Viewpoint Estimation using CNN features. IEEE International Conference on Vehicular Electronics and Safety (ICVES) 2017.
56 bin 57.13 % 63.05 % 50.64 % 15ms s GPU @ >3.5 Ghz (Python)
57 cas_retina 56.46 % 72.52 % 52.63 % 0.2 s 4 cores @ 2.5 Ghz (Python)
58 AVOD
This method makes use of Velodyne laser scans.
 code 56.01 % 65.72 % 48.89 % 0.08 s Titan X (pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
59 cas_retina_1_13 55.81 % 71.59 % 52.16 % 0.03 s 4 cores @ 2.5 Ghz (Python)
60 ReSqueeze 54.93 % 68.34 % 49.19 % 0.03 s GPU @ >3.5 Ghz (Python)
61 anm 50.54 % 67.40 % 45.22 % 3 s 1 core @ 2.5 Ghz (C/C++)
62 ZKNet 50.24 % 66.44 % 44.19 % 0.01 s GPU @ 2.0 Ghz (Python)
63 LPN 50.02 % 65.33 % 44.85 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
64 NEUAV 49.75 % 68.20 % 43.77 % 0.06 s GPU @ 2.5 Ghz (Python)
65 yolo800 49.15 % 64.64 % 43.58 % 0.13 s 4 cores @ 2.5 Ghz (Python)
66 BirdNet
This method makes use of Velodyne laser scans.
 49.04 % 64.88 % 46.61 % 0.11 s Titan Xp GPU
J. Beltrán, C. Guindel, F. Moreno, D. Cruzado, F. García and A. Escalera: BirdNet: A 3D Object Detection Framework from LiDAR Information. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
67 mylsi-faster-rcnn 48.85 % 67.93 % 43.77 % 0.3 s 1 core @ 2.5 Ghz (Python)
68 fasterrcnn 48.81 % 64.40 % 42.74 % 0.2 s 4 cores @ 2.5 Ghz (Python)
69 X_MD 48.07 % 63.46 % 40.76 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
70 detectron code 48.06 % 64.73 % 40.75 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
71 RFCN 47.61 % 62.17 % 43.74 % 0.2 s 4 cores @ 2.5 Ghz (Python)
72 CLF3D
This method makes use of Velodyne laser scans.
 47.53 % 65.31 % 40.23 % 0.13 s GPU @ 2.5 Ghz (Python)
73 NM code 47.20 % 60.64 % 42.96 % 0.01 s GPU @ 2.5 Ghz (Python)
74 RFCN_RFB 45.36 % 59.49 % 41.63 % 0.2 s 4 cores @ 2.5 Ghz (Python)
75 cascade_gw 45.00 % 63.14 % 38.81 % 0.2 s 4 cores @ 2.5 Ghz (Python)
76 myfaster-rcnn-v1.5 44.32 % 59.99 % 39.88 % 0.1 s 1 core @ 2.5 Ghz (Python)
77 FD2 44.29 % 62.32 % 40.65 % 0.01 s GPU @ >3.5 Ghz (Python + C/C++)
78 SeRC 44.28 % 55.81 % 38.50 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
79 Cmerge 43.85 % 61.60 % 42.60 % 0.2 s 4 cores @ 2.5 Ghz (Python)
80 Int-YOLO code 43.30 % 52.88 % 36.57 % 0.03 s 1 core @ 2.5 Ghz (C/C++)
ERROR: Wrong syntax in BIBTEX file.
81 MTDP 43.08 % 54.53 % 38.79 % 0.15 s GPU @ 2.0 Ghz (Python)
82 GNN 42.65 % 59.43 % 37.72 % 0.2 s 1 core @ 2.5 Ghz (Python)
83 MV-RGBD-RF
This method makes use of Velodyne laser scans.
 42.61 % 51.46 % 37.42 % 4 s 4 cores @ 2.5 Ghz (C/C++)
A. Gonzalez, D. Vazquez, A. Lopez and J. Amores: On-Board Object Detection: Multicue, Multimodal, and Multiview Random Forest of Local Experts.. IEEE Trans. on Cybernetics 2016.
A. Gonzalez, G. Villalonga, J. Xu, D. Vazquez, J. Amores and A. Lopez: Multiview Random Forest of Local Experts Combining RGB and LIDAR data for Pedestrian Detection. IEEE Intelligent Vehicles Symposium (IV) 2015.
84 YOLOv3+d 42.60 % 59.08 % 40.77 % 0.04 s GPU @ 1.5 Ghz (C/C++)
85 Shift R-CNN 42.30 % 65.56 % 41.40 % 0.25 s GPU @ 1.5 Ghz (Python)
A. Naiden, V. Paunescu, G. Kim, B. Jeon and M. Leordeanu: Shift R-CNN: Deep Monocular 3D Object Detection With Closed-form Geometric Constraints. ICIP 2019.
86 myfaster-rcnn 38.72 % 54.28 % 34.63 % 0.01 s 1 core @ 2.5 Ghz (Python)
87 SS3D 37.90 % 53.79 % 35.12 % 48 ms Tesla V100 (Python)
88 pAUCEnsT 37.88 % 52.28 % 33.38 % 60 s 1 core @ 2.5 Ghz (Matlab + C/C++)
S. Paisitkriangkrai, C. Shen and A. Hengel: Pedestrian Detection with Spatially Pooled Features and Structured Ensemble Learning. arXiv 2014.
89 Retinanet100 37.54 % 46.39 % 30.82 % 0.2 s 4 cores @ 2.5 Ghz (Python)
90 TopNet-Retina
This method makes use of Velodyne laser scans.
 35.20 % 50.28 % 34.11 % 52ms GeForce 1080Ti (tensorflow-gpu, v1.12)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
91 yolov3_warp 34.39 % 48.21 % 29.30 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
ERROR: Wrong syntax in BIBTEX file.
92 softyolo 31.30 % 45.16 % 27.38 % 0.16 s 4 cores @ 2.5 Ghz (Python)
93 Vote3D
This method makes use of Velodyne laser scans.
 31.24 % 41.45 % 28.60 % 0.5 s 4 cores @ 2.8 Ghz (C/C++)
D. Wang and I. Posner: Voting for Voting in Online Point Cloud Object Detection. Proceedings of Robotics: Science and Systems 2015.
94 DPM-VOC+VP 31.16 % 43.65 % 28.29 % 8 s 1 core @ 2.5 Ghz (C/C++)
B. Pepik, M. Stark, P. Gehler and B. Schiele: Multi-view and 3D Deformable Part Models. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2015.
95 LSVM-MDPM-us code 30.81 % 40.31 % 28.17 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
96 100Frcnn 29.95 % 44.60 % 27.70 % 2 s 4 cores @ 2.5 Ghz (Python + C/C++)
97 LSVM-MDPM-sv 29.24 % 37.71 % 27.52 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
A. Geiger, C. Wojek and R. Urtasun: Joint 3D Estimation of Objects and Scene Layout. NIPS 2011.
98 DPM-C8B1
This method uses stereo information.
 29.04 % 43.28 % 26.20 % 15 s 4 cores @ 2.5 Ghz (Matlab + C/C++)
J. Yebes, L. Bergasa and M. García-Garrido: Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes. Sensors 2015.
J. Yebes, L. Bergasa, R. Arroyo and A. Lázaro: Supervised learning and evaluation of KITTI's cars detector with DPM. IV 2014.
99 R-CNN_VGG 28.79 % 37.71 % 25.82 % 10 s GPU @ 2.5 Ghz (Matlab + C/C++)
100 rpn 28.65 % 37.40 % 23.50 % 0.01 s 1 core @ 2.5 Ghz (Python)
101 Lidar_ROI+Yolo(UJS) 27.21 % 39.41 % 26.12 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
102 mBoW
This method makes use of Velodyne laser scans.
 21.62 % 28.19 % 20.93 % 10 s 1 core @ 2.5 Ghz (C/C++)
J. Behley, V. Steinhage and A. Cremers: Laser-based Segment Classification Using a Mixture of Bag-of-Words. Proc. of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2013.
103 DT3D 20.65 % 31.29 % 20.73 % 0,21s GPU @ 2.5 Ghz (Python)
104 mymask-rcnn 19.58 % 23.22 % 18.87 % 0.3 s 1 core @ 2.5 Ghz (Python)
105 TopNet-HighRes
This method makes use of Velodyne laser scans.
 19.15 % 29.34 % 19.69 % 101ms NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
106 SA_3D 18.51 % 22.79 % 15.33 % 0.3 s GPU @ 2.5 Ghz (Python)
107 KD53-20 17.71 % 23.15 % 17.30 % 0.19 s 4 cores @ 2.5 Ghz (Python)
108 RT3DStereo
This method uses stereo information.
 17.66 % 23.73 % 11.94 % 0.08 s GPU @ 2.5 Ghz (C/C++)
109 TopNet-UncEst
This method makes use of Velodyne laser scans.
 16.21 % 19.18 % 15.99 % 0.09 s NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, M. Braun, M. Lauer and C. Stiller: Capturing Object Detection Uncertainty in Multi-Layer Grid Maps. 2019.
110 SA_3D
This method makes use of Velodyne laser scans.
 15.82 % 20.11 % 12.83 % 0.3 s 1 core @ 2.5 Ghz (Python)
111 YOLOv2 code 4.55 % 4.55 % 4.55 % 0.02 s GPU @ 3.5 Ghz (C/C++)
J. Redmon, S. Divvala, R. Girshick and A. Farhadi: You only look once: Unified, real-time object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016.
J. Redmon and A. Farhadi: YOLO9000: Better, Faster, Stronger. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2017.
112 TopNet-DecayRate
This method makes use of Velodyne laser scans.
 1.01 % 0.04 % 1.01 % 92 ms NVIDIA GeForce 1080 Ti (tensorflow-gpu)
S. Wirges, T. Fischer, C. Stiller and J. Frias: Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
113 softretina 0.44 % 0.29 % 0.22 % 0.16 s 4 cores @ 2.5 Ghz (Python)
114 JSyolo 0.22 % 0.22 % 0.22 % 0.16 s 4 cores @ 2.5 Ghz (Python)
Table as LaTeX | Only published Methods

Object Detection and Orientation Estimation Evaluation

Cars


Method Setting Code Moderate Easy Hard Runtime Environment
1 MVRA + I-FRCNN+ 89.93 % 90.60 % 79.78 % 0.18 s GPU @ 2.5 Ghz (Python)
2 Deep MANTA 89.86 % 97.19 % 80.39 % 0.7 s GPU @ 2.5 Ghz (Python + C/C++)
F. Chabot, M. Chaouch, J. Rabarisoa, C. Teulière and T. Chateau: Deep MANTA: A Coarse-to-fine Many-Task Network for joint 2D and 3D vehicle analysis from monocular image. CVPR 2017.
3 THICV-YDM 89.66 % 90.22 % 79.89 % 0.06 s GPU @ 2.5 Ghz (Python)
4 F-ConvNet
This method makes use of Velodyne laser scans.
 89.60 % 90.41 % 80.39 % 0.47 s GPU @ 2.5 Ghz (Python + C/C++)
Z. Wang and K. Jia: Frustum ConvNet: Sliding Frustums to Aggregate Local Point-Wise Features for Amodal 3D Object Detection. arXiv preprint arXiv:1903.01864 2019.
5 PointRCNN-deprecated
This method makes use of Velodyne laser scans.
 89.55 % 90.76 % 80.76 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
6 Patches
This method makes use of Velodyne laser scans.
 89.48 % 90.73 % 87.18 % 0.15 s GPU @ 2.0 Ghz
7 Fusion-RCNN 89.48 % 90.49 % 80.40 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
8 Patches - EMP
This method makes use of Velodyne laser scans.
 89.43 % 94.61 % 87.81 % 0.5 s GPU @ 2.5 Ghz (Python)
9 Part-A^2
This method makes use of Velodyne laser scans.
 89.38 % 90.61 % 87.93 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
10 PointRCNN
This method makes use of Velodyne laser scans.
 code 89.22 % 90.73 % 85.53 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
S. Shi, X. Wang and H. Li: PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud. CVPR 2019.
11 Alibaba-AILabsX
This method makes use of Velodyne laser scans.
 89.20 % 90.25 % 80.36 % 0.05 s 1 core @ >3.5 Ghz (C/C++)
12 SRF 89.00 % 90.33 % 80.36 % 0.05 s GPU @ 2.5 Ghz (Python + C/C++)
13 Voxel-FPN 88.92 % 90.23 % 80.02 % 0.02 s GPU @ 2.5 Ghz (Python + C/C++)
14 SegVoxelNet 88.88 % 90.50 % 87.34 % 0.04 s 1 core @ 2.5 Ghz (Python)
15 PFPN 88.83 % 90.30 % 79.99 % 0.02 s 4 cores @ >3.5 Ghz (Python)
16 PointPillars
This method makes use of Velodyne laser scans.
 code 88.76 % 90.19 % 86.38 % 16 ms 1080ti GPU and Intel i7 CPU
A. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang and O. Beijbom: PointPillars: Fast Encoders for Object Detection from Point Clouds. CVPR 2019.
17 MMV 88.74 % 90.38 % 79.99 % 0.4 s GPU @ 2.5 Ghz (C/C++)
18 RAL
This method makes use of Velodyne laser scans.
 88.72 % 90.08 % 80.06 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
19 Sogo_MM 88.72 % 90.67 % 78.95 % 1.5 s GPU @ 2.5 Ghz (C/C++)
20 MVSLN 88.62 % 90.52 % 80.08 % 0.1s s 1 core @ 2.5 Ghz (C/C++)
21 TBA 88.62 % 90.03 % 86.29 % 0.07 s 1 core @ 2.5 Ghz (Python)
22 Deep3DBox 88.56 % 90.39 % 77.17 % 1.5 s GPU @ 2.5 Ghz (C/C++)
A. Mousavian, D. Anguelov, J. Flynn and J. Kosecka: 3D Bounding Box Estimation Using Deep Learning and Geometry. CVPR 2017.
23 NU-optim 88.51 % 89.82 % 87.12 % 0.04 s GPU @ >3.5 Ghz (Python)
24 SubCNN 88.43 % 90.61 % 78.63 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
Y. Xiang, W. Choi, Y. Lin and S. Savarese: Subcategory-aware Convolutional Neural Networks for Object Proposals and Detection. IEEE Winter Conference on Applications of Computer Vision (WACV) 2017.
25 A-VoxelNet 88.36 % 89.72 % 79.61 % 0.029 s GPU @ 2.5 Ghz (Python)
26 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 88.33 % 90.05 % 84.80 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
27 SECOND-V1.5
This method makes use of Velodyne laser scans.
 code 88.33 % 90.15 % 79.65 % 0.04 s GPU @ 2.0 Ghz (Python + C/C++)
28 MPNet
This method makes use of Velodyne laser scans.
 88.27 % 90.27 % 85.06 % 0.02 s GPU @ 2.5 Ghz (Python + C/C++)
29 3DBN
This method makes use of Velodyne laser scans.
 87.95 % 89.93 % 79.32 % 0.13s 1080Ti (Python+C/C++)
X. Li, J. Guivant, N. Kwok and Y. Xu: 3D Backbone Network for 3D Object Detection. CoRR 2019.
30 Shift R-CNN 87.91 % 90.27 % 78.72 % 0.25 s GPU @ 1.5 Ghz (Python)
A. Naiden, V. Paunescu, G. Kim, B. Jeon and M. Leordeanu: Shift R-CNN: Deep Monocular 3D Object Detection With Closed-form Geometric Constraints. ICIP 2019.
31 MonoPSR 87.83 % 89.88 % 70.48 % 0.2 s GPU @ 3.5 Ghz (Python)
J. Ku*, A. Pon* and S. Waslander: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. CVPR 2019.
32 PAD 87.73 % 90.06 % 82.30 % 0.15 s 1 core @ 2.5 Ghz (Python)
33 ARPNET 87.68 % 89.79 % 79.05 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
34 CFR
This method makes use of Velodyne laser scans.
 87.67 % 90.26 % 79.02 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
35 PP_v1.0 code 87.61 % 89.97 % 83.02 % 0.02s 1 core @ 2.5 Ghz (C/C++)
36 AVOD
This method makes use of Velodyne laser scans.
 code 87.46 % 89.59 % 79.54 % 0.08 s Titan X (pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
37 SeoulRobotics-HFD
This method makes use of Velodyne laser scans.
 87.34 % 89.88 % 79.32 % 0.035 s GPU (C++)
38 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 87.13 % 89.95 % 79.74 % 0.1 s Titan X (Pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
39 DFD 87.01 % 89.72 % 78.98 % 0.05 s GPU @ 2.0 Ghz (Python + C/C++)
40 SECA 86.80 % 89.42 % 78.81 % 0.09 s GPU @ 2.5 Ghz (Python)
41 SCANet 86.65 % 89.06 % 78.67 % 0.09s GPU @ 2.5 Ghz (Python)
42 DeepStereoOP 86.57 % 89.01 % 77.13 % 3.4 s GPU @ 3.5 Ghz (Matlab + C/C++)
C. Pham and J. Jeon: Robust Object Proposals Re-ranking for Object Detection in Autonomous Driving Using Convolutional Neural Networks. Signal Processing: Image Communiation 2017.
43 SCANet 86.39 % 89.25 % 78.65 % 0.17 s >8 cores @ 2.5 Ghz (Python)
44 FQNet 86.29 % 89.48 % 74.40 % 0.5 s 1 core @ 2.5 Ghz (Python)
45 RAR-Net 86.28 % 89.48 % 74.39 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
46 ELLIOT
This method makes use of Velodyne laser scans.
 86.13 % 89.69 % 80.50 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
47 Mono3D code 85.83 % 89.00 % 76.00 % 4.2 s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Z. Zhang, H. Ma, S. Fidler and R. Urtasun: Monocular 3D Object Detection for Autonomous Driving. CVPR 2016.
48 3DOP
This method uses stereo information.
 code 85.81 % 88.56 % 76.21 % 3s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Y. Zhu, A. Berneshawi, H. Ma, S. Fidler and R. Urtasun: 3D Object Proposals for Accurate Object Class Detection. NIPS 2015.
49 MBR-SSD 85.03 % 88.10 % 75.92 % 4.0 s GPU @ 2.5 Ghz (Python + C/C++)
50 StereoFENet
This method uses stereo information.
 83.13 % 88.83 % 76.33 % 0.15 s 1 core @ 3.5 Ghz (Python)
W. Bao, B. Xu and Z. Chen: MonoFENet: Monocular 3D Object Detection with Feature Enhancement Networks. 2019.
51 MonoFENet 82.05 % 88.86 % 75.63 % 0.15 s 1 core @ 3.5 Ghz (Python)
W. Bao, B. Xu and Z. Chen: MonoFENet: Monocular 3D Object Detection with Feature Enhancement Networks. 2019.
52 SECOND code 81.31 % 87.84 % 71.95 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
53 X_MD 80.28 % 89.53 % 79.14 % 0.2 s 1 core @ 2.5 Ghz (Python + C/C++)
54 FNV1_Fusion 80.12 % 89.25 % 78.58 % 0.11 s GPU @ 2.5 Ghz (Python)
55 FNV1_RPN 80.10 % 89.27 % 78.66 % 0.12 s 1 core @ 2.5 Ghz (Python + C/C++)
56 SS3D 79.70 % 89.02 % 69.91 % 48 ms Tesla V100 (Python)
57 SECA 79.56 % 89.11 % 78.14 % 1 s GPU @ 2.5 Ghz (Python)
58 VSE 79.56 % 89.11 % 78.14 % 0.15 s GPU @ 2.5 Ghz (Python)
59 Complexer-YOLO
This method makes use of Velodyne laser scans.
 79.08 % 87.97 % 78.75 % 0.06 s GPU @ 3.5 Ghz (C/C++)
60 FNV1 78.97 % 88.40 % 76.70 % 0.11 s GPU @ 2.5 Ghz (Python)
61 BS3D 78.68 % 89.28 % 68.52 % 22 ms Titan Xp
62 Pseudo-LiDAR V2
This method uses stereo information.
 78.61 % 89.90 % 74.97 % 0.4 s GPU @ 2.5 Ghz (Python)
63 FRCNN+Or code 77.61 % 88.52 % 67.69 % 0.09 s Titan Xp GPU
C. Guindel, D. Martin and J. Armingol: Fast Joint Object Detection and Viewpoint Estimation for Traffic Scene Understanding. IEEE Intelligent Transportation Systems Magazine 2018.
C. Guindel, D. Martin and J. Armingol: Joint Object Detection and Viewpoint Estimation using CNN features. IEEE International Conference on Vehicular Electronics and Safety (ICVES) 2017.
64 Manhnet 77.21 % 85.58 % 60.50 % 26 ms 1 core @ 2.5 Ghz (C/C++)
65 CLF3D
This method makes use of Velodyne laser scans.
 76.50 % 84.35 % 67.12 % 0.13 s GPU @ 2.5 Ghz (Python)
66 avodC 76.30 % 86.31 % 68.71 % 0.1 s GPU @ 2.5 Ghz (Python)
67 3D FCN
This method makes use of Velodyne laser scans.
 75.71 % 85.46 % 68.19 % >5 s 1 core @ 2.5 Ghz (C/C++)
B. Li: 3D Fully Convolutional Network for Vehicle Detection in Point Cloud. IROS 2017.
68 3D-SSMFCNN code 75.42 % 75.44 % 67.27 % 0.1 s GPU @ 1.5 Ghz (C/C++)
L. Novak: Vehicle Detection and Pose Estimation for Autonomous Driving. 2017.
69 Pose-RCNN 75.35 % 88.78 % 61.47 % 2 s >8 cores @ 2.5 Ghz (Python)
M. Braun, Q. Rao, Y. Wang and F. Flohr: Pose-RCNN: Joint object detection and pose estimation using 3D object proposals. Intelligent Transportation Systems (ITSC), 2016 IEEE 19th International Conference on 2016.
70 GS3D 75.16 % 83.52 % 59.59 % 2 s 1 core @ 2.5 Ghz (C/C++)
B. Li, W. Ouyang, L. Sheng, X. Zeng and X. Wang: GS3D: An Efficient 3D Object Detection Framework for Autonomous Driving. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2019.
71 3DVP code 74.59 % 81.02 % 64.11 % 40 s 8 cores @ 3.5 Ghz (Matlab + C/C++)
Y. Xiang, W. Choi, Y. Lin and S. Savarese: Data-Driven 3D Voxel Patterns for Object Category Recognition. IEEE Conference on Computer Vision and Pattern Recognition 2015.
72 SubCat code 74.42 % 80.74 % 58.83 % 0.7 s 6 cores @ 3.5 Ghz (Matlab + C/C++)
E. Ohn-Bar and M. Trivedi: Learning to Detect Vehicles by Clustering Appearance Patterns. T-ITS 2015.
73 BdCost+DA+MS 73.15 % 82.12 % 58.29 % TBD s 4 cores @ 2.5 Ghz (Matlab/C++)
74 ROI-10D 67.85 % 74.24 % 59.28 % 0.2 s GPU @ 3.5 Ghz (Python)
F. Manhardt, W. Kehl and A. Gaidon: ROI-10D: Monocular Lifting of 2D Detection to 6D Pose and Metric Shape. Computer Vision and Pattern Recognition (CVPR) 2019.
75 MF3D 67.68 % 87.79 % 57.57 % 0.03 s GPU @ 2.5 Ghz (C/C++)
76 multi-task CNN 66.19 % 76.69 % 58.11 % 25.1 ms GPU @ 2.0 Ghz (Python)
M. Oeljeklaus, F. Hoffmann and T. Bertram: A Fast Multi-Task CNN for Spatial Understanding of Traffic Scenes. IEEE Intelligent Transportation Systems Conference 2018.
77 BdCost48LDCF code 66.01 % 77.10 % 50.35 % 0.5 s 8 cores @ 3.5 Ghz (Matlab + C/C++)
A. Fernández-Baldera, J. Buenaposada and L. Baumela: BAdaCost: Multi-class Boosting with Costs . Pattern Recognition 2018.
78 BdCost48-25C 65.25 % 77.59 % 50.68 % 4 s 1 core @ 2.5 Ghz (C/C++)
79 OC-DPM 64.88 % 74.66 % 52.24 % 10 s 8 cores @ 2.5 Ghz (Matlab)
B. Pepik, M. Stark, P. Gehler and B. Schiele: Occlusion Patterns for Object Class Detection. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2013.
80 3DVSSD 64.72 % 77.22 % 57.56 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
81 DPM-VOC+VP 63.27 % 77.51 % 47.57 % 8 s 1 core @ 2.5 Ghz (C/C++)
B. Pepik, M. Stark, P. Gehler and B. Schiele: Multi-view and 3D Deformable Part Models. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2015.
82 AOG-View 62.25 % 77.37 % 50.44 % 3 s 1 core @ 2.5 Ghz (Matlab, C/C++)
B. Li, T. Wu and S. Zhu: Integrating Context and Occlusion for Car Detection by Hierarchical And-Or Model. ECCV 2014.
83 monoref3d 57.65 % 73.74 % 48.91 % 0.1 s 1 core @ 2.5 Ghz (Python)
84 ref3D 57.65 % 73.74 % 48.91 % 0.1 s 1 core @ 2.5 Ghz (Python + C/C++)
85 SAMME48LDCF code 57.49 % 75.12 % 46.64 % 0.5 s 8 cores @ 3.5 Ghz (Matlab + C/C++)
A. Fernández-Baldera, J. Buenaposada and L. Baumela: BAdaCost: Multi-class Boosting with Costs . Pattern Recognition 2018.
86 LSVM-MDPM-sv 56.69 % 70.86 % 45.91 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
A. Geiger, C. Wojek and R. Urtasun: Joint 3D Estimation of Objects and Scene Layout. NIPS 2011.
87 ref3D 56.54 % 74.41 % 47.15 % 0.1 s 1 core @ 2.5 Ghz (Python)
88 RCN-resnet101 53.93 % 56.36 % 48.32 % 0.3 s GPU @ 2.5 Ghz (Python + C/C++)
89 SAG-Net 53.29 % 57.92 % 47.73 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
90 VeloFCN
This method makes use of Velodyne laser scans.
 52.70 % 70.21 % 46.11 % 1 s GPU @ 2.5 Ghz (Python + C/C++)
B. Li, T. Zhang and T. Xia: Vehicle Detection from 3D Lidar Using Fully Convolutional Network. RSS 2016 .
91 DPM-C8B1
This method uses stereo information.
 50.32 % 59.53 % 39.22 % 15 s 4 cores @ 2.5 Ghz (Matlab + C/C++)
J. Yebes, L. Bergasa and M. García-Garrido: Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes. Sensors 2015.
J. Yebes, L. Bergasa, R. Arroyo and A. Lázaro: Supervised learning and evaluation of KITTI's cars detector with DPM. IV 2014.
92 VAT-Net 49.91 % 52.74 % 45.16 % 1 s GPU @ 2.5 Ghz (Python + C/C++)
93 InNet 49.55 % 52.32 % 44.79 % 0.16 s GPU @ 3.5 Ghz (Python + C/C++)
94 ODES code 48.06 % 46.22 % 42.43 % 0.02 s GPU @ 2.5 Ghz (Python)
ERROR: Wrong syntax in BIBTEX file.
95 ReSqueeze 45.40 % 47.38 % 41.68 % 0.03 s GPU @ >3.5 Ghz (Python)
96 sensekitti code 44.56 % 47.06 % 41.50 % 4.5 s GPU @ 2.5 Ghz (Python + C/C++)
B. Yang, J. Yan, Z. Lei and S. Li: Craft Objects from Images. CVPR 2016.
97 VCTNet 43.41 % 46.53 % 39.42 % 0.02 s GPU @ 1.5 Ghz (C/C++)
98 Resnet101Faster rcnn 42.62 % 49.41 % 38.21 % 1 s 1 core @ 2.5 Ghz (Python)
99 FD2 39.44 % 47.56 % 35.20 % 0.01 s GPU @ >3.5 Ghz (Python + C/C++)
100 bin 37.23 % 41.94 % 32.65 % 15ms s GPU @ >3.5 Ghz (Python)
101 IPOD 37.01 % 36.95 % 36.96 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
102 cas+res+soft 36.06 % 36.91 % 31.97 % 0.2 s 4 cores @ 2.5 Ghz (Python)
103 cas_retina 36.05 % 38.05 % 31.81 % 0.2 s 4 cores @ 2.5 Ghz (Python)
104 merge12-12 36.02 % 36.92 % 31.88 % 0.2 s 4 cores @ 2.5 Ghz (Python)
105 BirdNet
This method makes use of Velodyne laser scans.
 35.81 % 50.85 % 34.90 % 0.11 s Titan Xp GPU
J. Beltrán, C. Guindel, F. Moreno, D. Cruzado, F. García and A. Escalera: BirdNet: A 3D Object Detection Framework from LiDAR Information. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
106 AtrousDet 35.80 % 36.56 % 32.32 % 0.05 s TITAN X
107 cas_retina_1_13 35.68 % 38.39 % 31.89 % 0.03 s 4 cores @ 2.5 Ghz (Python)
108 cascadercnn 35.01 % 34.13 % 28.55 % 0.36 s 4 cores @ 2.5 Ghz (Python)
109 IoU_DCRCNN 33.60 % 38.00 % 31.21 % 0.66 s GPU @ 2.5 Ghz (Python)
110 cascade_gw 33.49 % 32.78 % 28.18 % 0.2 s 4 cores @ 2.5 Ghz (Python)
111 Cmerge 32.95 % 36.87 % 29.14 % 0.2 s 4 cores @ 2.5 Ghz (Python)
112 ZKNet 32.91 % 37.16 % 28.94 % 0.01 s GPU @ 2.0 Ghz (Python)
113 softretina 32.90 % 37.63 % 28.73 % 0.16 s 4 cores @ 2.5 Ghz (Python)
114 Fast-SSD 32.90 % 40.88 % 29.21 % 0.06 s GTX650Ti
115 Retinanet100 32.87 % 37.54 % 28.69 % 0.2 s 4 cores @ 2.5 Ghz (Python)
116 LPN 32.41 % 33.97 % 29.15 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
117 RTL3D 32.26 % 32.32 % 29.71 % 0.02 s GPU @ 2.5 Ghz (Python)
118 NM code 32.25 % 36.53 % 28.20 % 0.01 s GPU @ 2.5 Ghz (Python)
119 SceneNet 32.02 % 36.62 % 28.46 % 0.03 s GPU @ 2.5 Ghz (C/C++)
120 detectron code 31.71 % 35.58 % 28.18 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
121 RFCN_RFB 31.47 % 33.70 % 27.02 % 0.2 s 4 cores @ 2.5 Ghz (Python)
122 MTDP 31.04 % 34.12 % 27.50 % 0.15 s GPU @ 2.0 Ghz (Python)
123 AOG code 30.81 % 34.05 % 24.86 % 3 s 4 cores @ 2.5 Ghz (Matlab)
T. Wu, B. Li and S. Zhu: Learning And-Or Models to Represent Context and Occlusion for Car Detection and Viewpoint Estimation. TPAMI 2016.
B. Li, T. Wu and S. Zhu: Integrating Context and Occlusion for Car Detection by Hierarchical And-Or Model. ECCV 2014.
124 yolo800 30.47 % 31.53 % 27.01 % 0.13 s 4 cores @ 2.5 Ghz (Python)
125 VoxelNet(Unofficial) 30.29 % 33.54 % 27.36 % 0.5 s GPU @ 2.0 Ghz (Python)
126 RFCN 30.29 % 33.30 % 25.44 % 0.2 s 4 cores @ 2.5 Ghz (Python)
127 fasterrcnn 28.13 % 29.83 % 24.76 % 0.2 s 4 cores @ 2.5 Ghz (Python)
128 SubCat48LDCF code 26.78 % 34.43 % 19.46 % 0.5 s 8 cores @ 3.5 Ghz (Matlab + C/C++)
A. Fernández-Baldera, J. Buenaposada and L. Baumela: BAdaCost: Multi-class Boosting with Costs . Pattern Recognition 2018.
129 RFBnet 26.39 % 32.45 % 23.97 % 0.2 s 4 cores @ 2.5 Ghz (Python)
130 Lidar_ROI+Yolo(UJS) 25.40 % 28.93 % 22.51 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
131 CSoR
This method makes use of Velodyne laser scans.
 code 25.38 % 34.43 % 21.95 % 3.5 s 4 cores @ >3.5 Ghz (Python + C/C++)
L. Plotkin: PyDriver: Entwicklung eines Frameworks für räumliche Detektion und Klassifikation von Objekten in Fahrzeugumgebung. 2015.
132 100Frcnn 25.26 % 34.82 % 21.73 % 2 s 4 cores @ 2.5 Ghz (Python + C/C++)
133 RT3DStereo
This method uses stereo information.
 22.24 % 25.54 % 19.33 % 0.08 s GPU @ 2.5 Ghz (C/C++)
134 DLMB
This method makes use of Velodyne laser scans.
This is an online method (no batch processing).
 21.69 % 25.31 % 18.75 % 0.03 s 8 cores @ 3.5 Ghz (C/C++)
135 RT3D
This method makes use of Velodyne laser scans.
 18.98 % 24.23 % 20.56 % 0.09 s GPU @ 1.8Ghz
Y. Zeng, Y. Hu, S. Liu, J. Ye, Y. Han, X. Li and N. Sun: RT3D: Real-Time 3-D Vehicle Detection in LiDAR Point Cloud for Autonomous Driving. IEEE Robotics and Automation Letters 2018.
136 softyolo 18.22 % 25.50 % 15.97 % 0.16 s 4 cores @ 2.5 Ghz (Python)
137 rpn 17.04 % 25.68 % 13.96 % 0.01 s 1 core @ 2.5 Ghz (Python)
138 Licar
This method makes use of Velodyne laser scans.
 15.58 % 18.24 % 16.15 % 0.09 s GPU @ 2.0 Ghz (Python)
139 KD53-20 14.27 % 20.79 % 12.61 % 0.19 s 4 cores @ 2.5 Ghz (Python)
140 DLnet 8.48 % 9.09 % 7.39 % 0.3 s 4 cores @ 2.5 Ghz (C/C++)
141 SN-net 0.00 % 0.00 % 0.00 % 0.8 s GPU @ 2.5 Ghz (Python + C/C++)
142 JSyolo 0.00 % 0.00 % 0.00 % 0.16 s 4 cores @ 2.5 Ghz (Python)
Table as LaTeX | Only published Methods


Pedestrians


Method Setting Code Moderate Easy Hard Runtime Environment
1 THICV-YDM 70.57 % 81.79 % 63.31 % 0.06 s GPU @ 2.5 Ghz (Python)
2 VMVS
This method makes use of Velodyne laser scans.
 67.66 % 78.57 % 63.83 % 0.25 s GPU @ 2.5 Ghz (Python)
3 Sogo_MM 66.83 % 78.89 % 62.06 % 1.5 s GPU @ 2.5 Ghz (C/C++)
4 SubCNN 66.28 % 78.33 % 61.37 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
Y. Xiang, W. Choi, Y. Lin and S. Savarese: Subcategory-aware Convolutional Neural Networks for Object Proposals and Detection. IEEE Winter Conference on Applications of Computer Vision (WACV) 2017.
5 F-ConvNet
This method makes use of Velodyne laser scans.
 64.32 % 72.73 % 59.07 % 0.47 s GPU @ 2.5 Ghz (Python + C/C++)
Z. Wang and K. Jia: Frustum ConvNet: Sliding Frustums to Aggregate Local Point-Wise Features for Amodal 3D Object Detection. arXiv preprint arXiv:1903.01864 2019.
6 Pose-RCNN 59.89 % 74.10 % 54.21 % 2 s >8 cores @ 2.5 Ghz (Python)
M. Braun, Q. Rao, Y. Wang and F. Flohr: Pose-RCNN: Joint object detection and pose estimation using 3D object proposals. Intelligent Transportation Systems (ITSC), 2016 IEEE 19th International Conference on 2016.
7 3DOP
This method uses stereo information.
 code 59.79 % 73.46 % 57.04 % 3s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Y. Zhu, A. Berneshawi, H. Ma, S. Fidler and R. Urtasun: 3D Object Proposals for Accurate Object Class Detection. NIPS 2015.
8 DeepStereoOP 59.28 % 73.37 % 56.87 % 3.4 s GPU @ 3.5 Ghz (Matlab + C/C++)
C. Pham and J. Jeon: Robust Object Proposals Re-ranking for Object Detection in Autonomous Driving Using Convolutional Neural Networks. Signal Processing: Image Communiation 2017.
9 Mono3D code 58.12 % 68.58 % 54.94 % 4.2 s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Z. Zhang, H. Ma, S. Fidler and R. Urtasun: Monocular 3D Object Detection for Autonomous Driving. CVPR 2016.
10 MonoPSR 56.30 % 70.56 % 49.84 % 0.2 s GPU @ 3.5 Ghz (Python)
J. Ku*, A. Pon* and S. Waslander: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. CVPR 2019.
11 FRCNN+Or code 52.62 % 66.84 % 48.72 % 0.09 s Titan Xp GPU
C. Guindel, D. Martin and J. Armingol: Fast Joint Object Detection and Viewpoint Estimation for Traffic Scene Understanding. IEEE Intelligent Transportation Systems Magazine 2018.
C. Guindel, D. Martin and J. Armingol: Joint Object Detection and Viewpoint Estimation using CNN features. IEEE International Conference on Vehicular Electronics and Safety (ICVES) 2017.
12 PointPillars
This method makes use of Velodyne laser scans.
 code 49.66 % 58.05 % 47.88 % 16 ms 1080ti GPU and Intel i7 CPU
A. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang and O. Beijbom: PointPillars: Fast Encoders for Object Detection from Point Clouds. CVPR 2019.
13 PointRCNN
This method makes use of Velodyne laser scans.
 code 48.98 % 57.49 % 46.48 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
S. Shi, X. Wang and H. Li: PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud. CVPR 2019.
14 Shift R-CNN 48.81 % 65.39 % 41.05 % 0.25 s GPU @ 1.5 Ghz (Python)
A. Naiden, V. Paunescu, G. Kim, B. Jeon and M. Leordeanu: Shift R-CNN: Deep Monocular 3D Object Detection With Closed-form Geometric Constraints. ICIP 2019.
15 CLF3D
This method makes use of Velodyne laser scans.
 46.86 % 62.19 % 44.92 % 0.13 s GPU @ 2.5 Ghz (Python)
16 SCANet 45.83 % 55.57 % 41.03 % 0.17 s >8 cores @ 2.5 Ghz (Python)
17 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 44.92 % 53.36 % 43.77 % 0.1 s Titan X (Pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
18 SECOND code 43.51 % 51.56 % 38.78 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
19 SS3D 43.45 % 52.70 % 37.20 % 48 ms Tesla V100 (Python)
20 CFR
This method makes use of Velodyne laser scans.
 43.34 % 56.83 % 42.44 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
21 ARPNET 42.40 % 54.90 % 41.21 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
22 DPM-VOC+VP 39.83 % 53.66 % 35.73 % 8 s 1 core @ 2.5 Ghz (C/C++)
B. Pepik, M. Stark, P. Gehler and B. Schiele: Multi-view and 3D Deformable Part Models. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2015.
23 VCTNet 38.60 % 43.57 % 36.82 % 0.02 s GPU @ 1.5 Ghz (C/C++)
24 HBA-RCNN 38.06 % 43.81 % 35.02 % 0.4 s 1 core @ 2.5 Ghz (C/C++)
25 sensekitti code 37.50 % 43.55 % 35.08 % 4.5 s GPU @ 2.5 Ghz (Python + C/C++)
B. Yang, J. Yan, Z. Lei and S. Li: Craft Objects from Images. CVPR 2016.
26 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 37.41 % 43.10 % 35.88 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
27 cas_retina_1_13 36.62 % 46.29 % 35.40 % 0.03 s 4 cores @ 2.5 Ghz (Python)
28 AVOD
This method makes use of Velodyne laser scans.
 code 36.38 % 44.12 % 31.81 % 0.08 s Titan X (pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
29 A-VoxelNet 36.27 % 41.58 % 35.14 % 0.029 s GPU @ 2.5 Ghz (Python)
30 AtrousDet 36.10 % 42.90 % 32.09 % 0.05 s TITAN X
31 LSVM-MDPM-sv 35.49 % 47.00 % 32.42 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
A. Geiger, C. Wojek and R. Urtasun: Joint 3D Estimation of Objects and Scene Layout. NIPS 2011.
32 IPOD 35.32 % 41.46 % 31.59 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
33 PP_v1.0 code 34.25 % 39.84 % 32.86 % 0.02s 1 core @ 2.5 Ghz (C/C++)
34 SubCat 34.18 % 43.95 % 30.76 % 1.2 s 6 cores @ 2.5 Ghz (Matlab + C/C++)
E. Ohn-Bar and M. Trivedi: Fast and Robust Object Detection Using Visual Subcategories. Computer Vision and Pattern Recognition Workshops Mobile Vision 2014.
35 CHTTL MMF 34.17 % 43.98 % 30.89 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
36 ELLIOT
This method makes use of Velodyne laser scans.
 34.11 % 41.90 % 32.18 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
37 cascadercnn 33.27 % 43.05 % 28.88 % 0.36 s 4 cores @ 2.5 Ghz (Python)
38 cas_retina 33.02 % 42.79 % 31.91 % 0.2 s 4 cores @ 2.5 Ghz (Python)
39 merge12-12 32.94 % 42.47 % 31.87 % 0.2 s 4 cores @ 2.5 Ghz (Python)
40 cas+res+soft 32.84 % 42.36 % 31.66 % 0.2 s 4 cores @ 2.5 Ghz (Python)
41 RPN+BF code 32.55 % 40.97 % 29.52 % 0.6 s GPU @ 2.5 Ghz (Matlab + C/C++)
L. Zhang, L. Lin, X. Liang and K. He: Is Faster R-CNN Doing Well for Pedestrian Detection?. ECCV 2016.
42 RFCN 32.48 % 40.51 % 28.66 % 0.2 s 4 cores @ 2.5 Ghz (Python)
43 X_MD 32.45 % 43.55 % 31.29 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
44 ReSqueeze 32.35 % 37.95 % 30.38 % 0.03 s GPU @ >3.5 Ghz (Python)
45 bin 31.81 % 36.25 % 29.83 % 15ms s GPU @ >3.5 Ghz (Python)
46 Complexer-YOLO
This method makes use of Velodyne laser scans.
 31.80 % 37.80 % 31.26 % 0.06 s GPU @ 3.5 Ghz (C/C++)
47 LPN 31.63 % 38.40 % 28.90 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
48 ZKNet 31.58 % 39.11 % 28.78 % 0.01 s GPU @ 2.0 Ghz (Python)
49 ODES code 31.43 % 36.84 % 29.00 % 0.02 s GPU @ 2.5 Ghz (Python)
ERROR: Wrong syntax in BIBTEX file.
50 yolo800 31.42 % 40.42 % 30.50 % 0.13 s 4 cores @ 2.5 Ghz (Python)
51 detectron code 31.20 % 41.08 % 30.78 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
52 RFCN_RFB 30.38 % 38.00 % 26.72 % 0.2 s 4 cores @ 2.5 Ghz (Python)
53 fasterrcnn 29.82 % 38.69 % 28.45 % 0.2 s 4 cores @ 2.5 Ghz (Python)
54 NM code 29.74 % 38.40 % 27.97 % 0.01 s GPU @ 2.5 Ghz (Python)
55 MTDP 29.04 % 36.90 % 25.96 % 0.15 s GPU @ 2.0 Ghz (Python)
56 FD2 28.59 % 35.53 % 26.02 % 0.01 s GPU @ >3.5 Ghz (Python + C/C++)
57 ACF 28.46 % 35.69 % 26.18 % 1 s 1 core @ 3.5 Ghz (Matlab + C/C++)
P. Doll\'ar, R. Appel, S. Belongie and P. Perona: Fast Feature Pyramids for Object Detection. PAMI 2014.
58 cascade_gw 27.51 % 36.55 % 23.20 % 0.2 s 4 cores @ 2.5 Ghz (Python)
59 multi-task CNN 26.98 % 33.58 % 23.07 % 25.1 ms GPU @ 2.0 Ghz (Python)
M. Oeljeklaus, F. Hoffmann and T. Bertram: A Fast Multi-Task CNN for Spatial Understanding of Traffic Scenes. IEEE Intelligent Transportation Systems Conference 2018.
60 softyolo 26.04 % 34.86 % 25.28 % 0.16 s 4 cores @ 2.5 Ghz (Python)
61 Cmerge 24.19 % 34.08 % 24.04 % 0.2 s 4 cores @ 2.5 Ghz (Python)
62 Resnet101Faster rcnn 23.89 % 30.25 % 23.38 % 1 s 1 core @ 2.5 Ghz (Python)
63 Lidar_ROI+Yolo(UJS) 23.43 % 28.50 % 19.87 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
64 DPM-C8B1
This method uses stereo information.
 23.37 % 31.08 % 20.72 % 15 s 4 cores @ 2.5 Ghz (Matlab + C/C++)
J. Yebes, L. Bergasa and M. García-Garrido: Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes. Sensors 2015.
J. Yebes, L. Bergasa, R. Arroyo and A. Lázaro: Supervised learning and evaluation of KITTI's cars detector with DPM. IV 2014.
65 Retinanet100 23.23 % 28.72 % 19.00 % 0.2 s 4 cores @ 2.5 Ghz (Python)
66 ACF-MR 23.18 % 29.35 % 21.00 % 0.6 s 1 core @ 3.5 Ghz (C/C++)
R. Rajaram, E. Ohn-Bar and M. Trivedi: Looking at Pedestrians at Different Scales: A Multi-resolution Approach and Evaluations. T-ITS 2016.
67 KD53-20 22.24 % 26.50 % 19.80 % 0.19 s 4 cores @ 2.5 Ghz (Python)
68 rpn 22.07 % 30.16 % 21.44 % 0.01 s 1 core @ 2.5 Ghz (Python)
69 RT3DStereo
This method uses stereo information.
 19.84 % 23.20 % 19.57 % 0.08 s GPU @ 2.5 Ghz (C/C++)
70 100Frcnn 18.55 % 23.61 % 18.34 % 2 s 4 cores @ 2.5 Ghz (Python + C/C++)
71 BirdNet
This method makes use of Velodyne laser scans.
 17.26 % 21.34 % 16.67 % 0.11 s Titan Xp GPU
J. Beltrán, C. Guindel, F. Moreno, D. Cruzado, F. García and A. Escalera: BirdNet: A 3D Object Detection Framework from LiDAR Information. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
72 softretina 0.49 % 0.35 % 0.50 % 0.16 s 4 cores @ 2.5 Ghz (Python)
73 JSyolo 0.23 % 0.20 % 0.25 % 0.16 s 4 cores @ 2.5 Ghz (Python)
74 SN-net 0.00 % 0.00 % 0.00 % 0.8 s GPU @ 2.5 Ghz (Python + C/C++)
Table as LaTeX | Only published Methods


Cyclists


Method Setting Code Moderate Easy Hard Runtime Environment
1 F-ConvNet
This method makes use of Velodyne laser scans.
 74.96 % 84.38 % 66.45 % 0.47 s GPU @ 2.5 Ghz (Python + C/C++)
Z. Wang and K. Jia: Frustum ConvNet: Sliding Frustums to Aggregate Local Point-Wise Features for Amodal 3D Object Detection. arXiv preprint arXiv:1903.01864 2019.
2 PointRCNN
This method makes use of Velodyne laser scans.
 code 72.35 % 83.40 % 65.50 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
S. Shi, X. Wang and H. Li: PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud. CVPR 2019.
3 ARPNET 69.13 % 79.33 % 62.24 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
4 PointPillars
This method makes use of Velodyne laser scans.
 code 68.16 % 82.43 % 61.96 % 16 ms 1080ti GPU and Intel i7 CPU
A. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang and O. Beijbom: PointPillars: Fast Encoders for Object Detection from Point Clouds. CVPR 2019.
5 A-VoxelNet 66.02 % 80.19 % 59.21 % 0.029 s GPU @ 2.5 Ghz (Python)
6 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 65.60 % 80.04 % 59.60 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
7 Sogo_MM 63.59 % 70.70 % 56.15 % 1.5 s GPU @ 2.5 Ghz (C/C++)
8 SubCNN 63.41 % 71.39 % 56.34 % 2 s GPU @ 3.5 Ghz (Python + C/C++)
Y. Xiang, W. Choi, Y. Lin and S. Savarese: Subcategory-aware Convolutional Neural Networks for Object Proposals and Detection. IEEE Winter Conference on Applications of Computer Vision (WACV) 2017.
9 Pose-RCNN 62.25 % 74.85 % 55.09 % 2 s >8 cores @ 2.5 Ghz (Python)
M. Braun, Q. Rao, Y. Wang and F. Flohr: Pose-RCNN: Joint object detection and pose estimation using 3D object proposals. Intelligent Transportation Systems (ITSC), 2016 IEEE 19th International Conference on 2016.
10 SCANet 61.96 % 72.59 % 55.26 % 0.17 s >8 cores @ 2.5 Ghz (Python)
11 Deep3DBox 59.37 % 68.58 % 51.97 % 1.5 s GPU @ 2.5 Ghz (C/C++)
A. Mousavian, D. Anguelov, J. Flynn and J. Kosecka: 3D Bounding Box Estimation Using Deep Learning and Geometry. CVPR 2017.
12 3DOP
This method uses stereo information.
 code 58.59 % 71.95 % 52.35 % 3s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Y. Zhu, A. Berneshawi, H. Ma, S. Fidler and R. Urtasun: 3D Object Proposals for Accurate Object Class Detection. NIPS 2015.
13 PP_v1.0 code 58.34 % 74.19 % 52.29 % 0.02s 1 core @ 2.5 Ghz (C/C++)
14 CFR
This method makes use of Velodyne laser scans.
 57.83 % 74.57 % 55.63 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
15 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 57.53 % 67.61 % 54.16 % 0.1 s Titan X (Pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
16 SECOND code 57.20 % 80.97 % 55.14 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
17 ELLIOT
This method makes use of Velodyne laser scans.
 56.42 % 74.07 % 51.61 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
18 Complexer-YOLO
This method makes use of Velodyne laser scans.
 56.32 % 64.51 % 56.23 % 0.06 s GPU @ 3.5 Ghz (C/C++)
19 DeepStereoOP 55.62 % 67.49 % 48.85 % 3.4 s GPU @ 3.5 Ghz (Matlab + C/C++)
C. Pham and J. Jeon: Robust Object Proposals Re-ranking for Object Detection in Autonomous Driving Using Convolutional Neural Networks. Signal Processing: Image Communiation 2017.
20 AVOD
This method makes use of Velodyne laser scans.
 code 54.43 % 64.36 % 47.67 % 0.08 s Titan X (pascal)
J. Ku, M. Mozifian, J. Lee, A. Harakeh and S. Waslander: Joint 3D Proposal Generation and Object Detection from View Aggregation. IROS 2018.
21 Mono3D code 53.11 % 65.74 % 48.87 % 4.2 s GPU @ 2.5 Ghz (Matlab + C/C++)
X. Chen, K. Kundu, Z. Zhang, H. Ma, S. Fidler and R. Urtasun: Monocular 3D Object Detection for Autonomous Driving. CVPR 2016.
22 FRCNN+Or code 50.91 % 63.41 % 45.46 % 0.09 s Titan Xp GPU
C. Guindel, D. Martin and J. Armingol: Fast Joint Object Detection and Viewpoint Estimation for Traffic Scene Understanding. IEEE Intelligent Transportation Systems Magazine 2018.
C. Guindel, D. Martin and J. Armingol: Joint Object Detection and Viewpoint Estimation using CNN features. IEEE International Conference on Vehicular Electronics and Safety (ICVES) 2017.
23 MonoPSR 49.30 % 58.93 % 43.45 % 0.2 s GPU @ 3.5 Ghz (Python)
J. Ku*, A. Pon* and S. Waslander: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. CVPR 2019.
24 CLF3D
This method makes use of Velodyne laser scans.
 46.66 % 64.55 % 39.30 % 0.13 s GPU @ 2.5 Ghz (Python)
25 X_MD 45.90 % 61.86 % 39.14 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
26 sensekitti code 42.12 % 46.65 % 36.66 % 4.5 s GPU @ 2.5 Ghz (Python + C/C++)
B. Yang, J. Yan, Z. Lei and S. Li: Craft Objects from Images. CVPR 2016.
27 VCTNet 37.64 % 46.21 % 33.46 % 0.02 s GPU @ 1.5 Ghz (C/C++)
28 Shift R-CNN 34.77 % 54.31 % 34.04 % 0.25 s GPU @ 1.5 Ghz (Python)
A. Naiden, V. Paunescu, G. Kim, B. Jeon and M. Leordeanu: Shift R-CNN: Deep Monocular 3D Object Detection With Closed-form Geometric Constraints. ICIP 2019.
29 ODES code 33.74 % 37.75 % 30.34 % 0.02 s GPU @ 2.5 Ghz (Python)
ERROR: Wrong syntax in BIBTEX file.
30 SS3D 31.17 % 44.77 % 28.96 % 48 ms Tesla V100 (Python)
31 BirdNet
This method makes use of Velodyne laser scans.
 30.76 % 41.48 % 28.66 % 0.11 s Titan Xp GPU
J. Beltrán, C. Guindel, F. Moreno, D. Cruzado, F. García and A. Escalera: BirdNet: A 3D Object Detection Framework from LiDAR Information. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) 2018.
32 bin 29.53 % 34.66 % 25.98 % 15ms s GPU @ >3.5 Ghz (Python)
33 IPOD 28.88 % 36.04 % 25.66 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
34 AtrousDet 28.47 % 33.27 % 25.62 % 0.05 s TITAN X
35 ReSqueeze 27.40 % 35.39 % 24.32 % 0.03 s GPU @ >3.5 Ghz (Python)
36 LPN 27.01 % 32.96 % 25.01 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
37 merge12-12 26.91 % 32.25 % 23.72 % 0.2 s 4 cores @ 2.5 Ghz (Python)
38 cas+res+soft 26.83 % 32.33 % 23.63 % 0.2 s 4 cores @ 2.5 Ghz (Python)
39 cascadercnn 26.62 % 33.02 % 23.01 % 0.36 s 4 cores @ 2.5 Ghz (Python)
40 detectron code 26.36 % 27.44 % 23.20 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
41 FD2 24.65 % 35.58 % 21.97 % 0.01 s GPU @ >3.5 Ghz (Python + C/C++)
42 cas_retina 24.58 % 30.82 % 23.79 % 0.2 s 4 cores @ 2.5 Ghz (Python)
43 cas_retina_1_13 24.37 % 30.31 % 25.81 % 0.03 s 4 cores @ 2.5 Ghz (Python)
44 DPM-VOC+VP 23.22 % 31.24 % 21.62 % 8 s 1 core @ 2.5 Ghz (C/C++)
B. Pepik, M. Stark, P. Gehler and B. Schiele: Multi-view and 3D Deformable Part Models. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2015.
45 LSVM-MDPM-sv 23.14 % 28.89 % 22.28 % 10 s 4 cores @ 3.0 Ghz (C/C++)
P. Felzenszwalb, R. Girshick, D. McAllester and D. Ramanan: Object Detection with Discriminatively Trained Part-Based Models. PAMI 2010.
A. Geiger, C. Wojek and R. Urtasun: Joint 3D Estimation of Objects and Scene Layout. NIPS 2011.
46 NM code 22.22 % 27.72 % 20.40 % 0.01 s GPU @ 2.5 Ghz (Python)
47 fasterrcnn 22.08 % 28.60 % 19.31 % 0.2 s 4 cores @ 2.5 Ghz (Python)
48 yolo800 21.69 % 28.20 % 19.53 % 0.13 s 4 cores @ 2.5 Ghz (Python)
49 ZKNet 21.50 % 28.20 % 19.12 % 0.01 s GPU @ 2.0 Ghz (Python)
50 RFCN 20.80 % 26.55 % 19.22 % 0.2 s 4 cores @ 2.5 Ghz (Python)
51 RFCN_RFB 20.44 % 25.95 % 18.78 % 0.2 s 4 cores @ 2.5 Ghz (Python)
52 cascade_gw 19.56 % 26.76 % 17.09 % 0.2 s 4 cores @ 2.5 Ghz (Python)
53 DPM-C8B1
This method uses stereo information.
 19.25 % 27.16 % 17.95 % 15 s 4 cores @ 2.5 Ghz (Matlab + C/C++)
J. Yebes, L. Bergasa and M. García-Garrido: Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes. Sensors 2015.
J. Yebes, L. Bergasa, R. Arroyo and A. Lázaro: Supervised learning and evaluation of KITTI's cars detector with DPM. IV 2014.
54 Cmerge 19.19 % 26.37 % 18.56 % 0.2 s 4 cores @ 2.5 Ghz (Python)
55 MTDP 18.95 % 23.33 % 17.24 % 0.15 s GPU @ 2.0 Ghz (Python)
56 Retinanet100 15.16 % 18.64 % 12.49 % 0.2 s 4 cores @ 2.5 Ghz (Python)
57 softyolo 12.14 % 16.84 % 10.51 % 0.16 s 4 cores @ 2.5 Ghz (Python)
58 100Frcnn 11.79 % 17.33 % 10.99 % 2 s 4 cores @ 2.5 Ghz (Python + C/C++)
59 rpn 11.30 % 14.62 % 8.94 % 0.01 s 1 core @ 2.5 Ghz (Python)
60 Lidar_ROI+Yolo(UJS) 9.31 % 13.88 % 9.12 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
61 KD53-20 6.15 % 7.81 % 6.35 % 0.19 s 4 cores @ 2.5 Ghz (Python)
62 RT3DStereo
This method uses stereo information.
 5.26 % 6.60 % 3.68 % 0.08 s GPU @ 2.5 Ghz (C/C++)
63 softretina 0.20 % 0.14 % 0.10 % 0.16 s 4 cores @ 2.5 Ghz (Python)
64 JSyolo 0.15 % 0.15 % 0.15 % 0.16 s 4 cores @ 2.5 Ghz (Python)
Table as LaTeX | Only published Methods


Related Datasets

Citation

When using this dataset in your research, we will be happy if you cite us:
@INPROCEEDINGS{Geiger2012CVPR,
  author = {Andreas Geiger and Philip Lenz and Raquel Urtasun},
  title = {Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite},
  booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2012}
}



eXTReMe Tracker