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

3D Object Detection Evaluation 2017


The 3D object detection benchmark consists of 7481 training images and 7518 test images as well as the corresponding point clouds, comprising a total of 80.256 labeled objects. For evaluation, we compute precision-recall curves. To rank the methods we compute average precision. 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 3D object detection performance using the PASCAL criteria also used for 2D object detection. Far objects are thus filtered based on their bounding box height in the image plane. As only objects also appearing on the image plane are labeled, objects in don't car areas do not count as false positives. We note that the evaluation does not take care of ignoring detections that are not visible on the image plane — these detections might give rise to false positives. For cars we require an 3D bounding box overlap of 70%, while for pedestrians and cyclists we require a 3D bounding box overlap of 50%. 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.

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 MMLab-PartA^2
This method makes use of Velodyne laser scans.
 77.86 % 85.94 % 72.00 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
S. Shi, Z. Wang, X. Wang and H. Li: Part-A^2 Net: 3D Part-Aware and Aggregation Neural Network for Object Detection from Point Cloud. arXiv preprint arXiv:1907.03670 2019.
2 Fusion-RCNN 77.84 % 86.60 % 69.15 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
3 STD 77.63 % 86.61 % 76.06 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
4 Patches - EMP
This method makes use of Velodyne laser scans.
 77.20 % 87.85 % 72.78 % 0.5 s GPU @ 2.5 Ghz (Python)
5 Patches
This method makes use of Velodyne laser scans.
 77.16 % 87.87 % 68.91 % 0.15 s GPU @ 2.0 Ghz
6 UberATG-MMF
This method makes use of Velodyne laser scans.
 76.75 % 86.81 % 68.41 % 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.
7 F-ConvNet
This method makes use of Velodyne laser scans.
 76.51 % 85.88 % 68.08 % 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. IROS 2019.
8 RAL
This method makes use of Velodyne laser scans.
 76.28 % 84.43 % 68.22 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
9 SRF 76.25 % 85.09 % 68.10 % 0.05 s GPU @ 2.5 Ghz (Python + C/C++)
10 Voxel-FPN 76.14 % 85.48 % 68.05 % 0.02 s GPU @ 2.5 Ghz (Python + C/C++)
11 RGB3D
This method makes use of Velodyne laser scans.
 75.92 % 85.72 % 68.29 % 0.39 s GPU @ 2.5 Ghz (Python)
12 SegVoxelNet 75.81 % 84.19 % 67.80 % 0.04 s 1 core @ 2.5 Ghz (Python)
13 epBRM
This method makes use of Velodyne laser scans.
 75.79 % 83.95 % 67.88 % 0.1 s GPU @ >3.5 Ghz (Python + C/C++)
14 MMLab-PointRCNN
This method makes use of Velodyne laser scans.
 code 75.76 % 85.94 % 68.32 % 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. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2019.
15 PFPN 75.74 % 85.66 % 67.68 % 0.02 s 4 cores @ >3.5 Ghz (Python)
16 Fast Point R-CNNv1.1
This method makes use of Velodyne laser scans.
 75.73 % 84.28 % 67.39 % 0.06 s GPU @ 2.5 Ghz (Python + C/C++)
17 Alibaba-AILabsX
This method makes use of Velodyne laser scans.
 75.69 % 84.59 % 67.80 % 0.05 s 1 core @ >3.5 Ghz (C/C++)
18 PTS
This method makes use of Velodyne laser scans.
 code 75.67 % 83.95 % 67.71 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
19 PointRCNN-deprecated
This method makes use of Velodyne laser scans.
 75.42 % 84.32 % 67.86 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
20 SECOND-V1.5
This method makes use of Velodyne laser scans.
 code 75.38 % 84.04 % 67.36 % 0.04 s GPU @ 2.0 Ghz (Python + C/C++)
21 MMV 75.35 % 84.21 % 67.36 % 0.4 s GPU @ 2.5 Ghz (C/C++)
22 TBA 75.18 % 79.90 % 68.48 % 0.07 s 1 core @ 2.5 Ghz (Python)
23 mypointrcnn 75.12 % 85.01 % 68.09 % 1 s 1 core @ 2.5 Ghz (Python)
24 ARPNET 75.03 % 84.44 % 67.37 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
25 PointPillars
This method makes use of Velodyne laser scans.
 code 74.99 % 79.05 % 68.30 % 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.
26 NU-optim 74.94 % 83.92 % 66.88 % 0.04 s GPU @ >3.5 Ghz (Python)
27 MPNet
This method makes use of Velodyne laser scans.
 74.92 % 83.84 % 67.30 % 0.02 s GPU @ 2.5 Ghz (Python + C/C++)
28 DH-ARI 74.90 % 80.02 % 68.46 % 0.2 s 1 core @ >3.5 Ghz (Python + C/C++)
29 Alibaba-AILabsX
This method makes use of Velodyne laser scans.
 74.87 % 84.02 % 67.55 % 0.2 s GPU @ >3.5 Ghz (Python)
30 A-VoxelNet 74.84 % 84.10 % 66.92 % 0.029 s GPU @ 2.5 Ghz (Python)
31 MVSLN 74.76 % 84.61 % 67.44 % 0.1s s 1 core @ 2.5 Ghz (C/C++)
32 3DBN
This method makes use of Velodyne laser scans.
 74.64 % 83.56 % 66.76 % 0.13s 1080Ti (Python+C/C++)
X. Li, J. Guivant, N. Kwok and Y. Xu: 3D Backbone Network for 3D Object Detection. CoRR 2019.
33 CFR
This method makes use of Velodyne laser scans.
 74.49 % 84.25 % 66.41 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
34 FOFNet
This method makes use of Velodyne laser scans.
 74.45 % 84.15 % 66.97 % 0.04 s GPU @ 2.5 Ghz (Python)
35 Fast Point R-CNN
This method makes use of Velodyne laser scans.
 74.43 % 83.45 % 66.38 % 0.06 s GPU @ 2.5 Ghz (Python + C/C++)
36 PC-CNN-V2
This method makes use of Velodyne laser scans.
 73.80 % 84.33 % 64.83 % 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.
37 AILabs3D
This method makes use of Velodyne laser scans.
 73.70 % 83.32 % 65.77 % 0.6 s GPU @ >3.5 Ghz (Python)
38 SECOND code 73.66 % 83.13 % 66.20 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
39 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 73.03 % 79.61 % 65.98 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
40 MVX-Net
This method makes use of Velodyne laser scans.
 72.67 % 83.19 % 65.22 % 0.06 s GPU @ 3.0 Ghz (Python + C/C++)
41 MDC
This method makes use of Velodyne laser scans.
 72.67 % 82.07 % 64.60 % 0.17 s GPU @ 2.5 Ghz (Python)
42 IPOD 72.57 % 79.75 % 66.33 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
43 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 71.88 % 81.94 % 66.38 % 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.
44 CONV-BOX
This method makes use of Velodyne laser scans.
 70.47 % 79.98 % 64.49 % 0.2 s Tesla V100
45 F-PointNet
This method makes use of Velodyne laser scans.
 code 70.39 % 81.20 % 62.19 % 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.
46 PAD 68.33 % 76.72 % 65.49 % 0.15 s 1 core @ 2.5 Ghz (Python)
47 PP_v1.0 code 68.12 % 77.99 % 65.34 % 0.02s 1 core @ 2.5 Ghz (C/C++)
48 SeoulRobotics-HFD
This method makes use of Velodyne laser scans.
 66.98 % 76.09 % 64.92 % 0.035 s GPU (C++)
49 ELLIOT
This method makes use of Velodyne laser scans.
 66.86 % 76.38 % 64.47 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
50 DFD 66.56 % 76.36 % 64.11 % 0.05 s GPU @ 2.0 Ghz (Python + C/C++)
51 SCANet 66.30 % 76.09 % 58.68 % 0.09s GPU @ 2.5 Ghz (Python)
52 UberATG-ContFuse
This method makes use of Velodyne laser scans.
 66.22 % 82.54 % 64.04 % 0.06 s GPU @ 2.5 Ghz (Python)
M. Liang, B. Yang, S. Wang and R. Urtasun: Deep Continuous Fusion for Multi-Sensor 3D Object Detection. ECCV 2018.
53 SCANet 65.99 % 75.66 % 63.48 % 0.17 s >8 cores @ 2.5 Ghz (Python)
54 AVOD
This method makes use of Velodyne laser scans.
 code 65.78 % 73.59 % 58.38 % 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.
55 SECA 65.75 % 75.98 % 58.44 % 1 s GPU @ 2.5 Ghz (Python)
56 VSE 65.75 % 75.98 % 58.44 % 0.15 s GPU @ 2.5 Ghz (Python)
57 RTL3D 65.72 % 80.42 % 63.50 % 0.02 s GPU @ 2.5 Ghz (Python)
58 Multi-3D
This method makes use of Velodyne laser scans.
 65.33 % 76.57 % 56.11 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
59 FNV1_RPN 65.18 % 74.61 % 57.75 % 0.12 s 1 core @ 2.5 Ghz (Python + C/C++)
60 FNV1_Fusion 65.07 % 74.78 % 57.74 % 0.11 s GPU @ 2.5 Ghz (Python)
61 X_MD 64.82 % 74.06 % 57.49 % 0.2 s 1 core @ 2.5 Ghz (Python + C/C++)
62 VoxelNet(Unofficial) 64.80 % 74.59 % 57.38 % 0.5 s GPU @ 2.0 Ghz (Python)
63 SECA 64.59 % 73.70 % 57.21 % 0.09 s GPU @ 2.5 Ghz (Python)
64 NLK 63.99 % 73.81 % 60.90 % 0.02 s 1 core @ 2.5 Ghz (Python)
65 MV3D
This method makes use of Velodyne laser scans.
 62.35 % 71.09 % 55.12 % 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.
66 FNV1 61.69 % 71.93 % 55.41 % 0.11 s GPU @ 2.5 Ghz (Python)
67 FNV2 59.26 % 67.67 % 51.97 % 0.18 s GPU @ 2.5 Ghz (Python)
68 CLF3D
This method makes use of Velodyne laser scans.
 58.48 % 65.54 % 46.54 % 0.13 s GPU @ 2.5 Ghz (Python)
69 A3DODWTDA
This method makes use of Velodyne laser scans.
 code 56.81 % 59.35 % 50.51 % 0.08 s GPU @ 3.0 Ghz (Python)
F. Gustafsson and E. Linder-Norén: Automotive 3D Object Detection Without Target Domain Annotations. 2018.
70 anm 56.76 % 68.02 % 49.39 % 3 s 1 core @ 2.5 Ghz (C/C++)
71 avodC 55.47 % 65.71 % 48.74 % 0.1 s GPU @ 2.5 Ghz (Python)
72 MV3D (LIDAR)
This method makes use of Velodyne laser scans.
 52.73 % 66.77 % 51.31 % 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.
73 Complexer-YOLO
This method makes use of Velodyne laser scans.
 49.44 % 55.63 % 44.13 % 0.06 s GPU @ 3.5 Ghz (C/C++)
M. Simon, K. Amende, A. Kraus, J. Honer, T. Samann, H. Kaulbersch, S. Milz and H. Michael Gross: Complexer-YOLO: Real-Time 3D Object Detection and Tracking on Semantic Point Clouds. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops 2019.
74 Pseudo-LiDAR V2
This method uses stereo information.
 code 44.56 % 60.41 % 38.52 % 0.4 s GPU @ 2.5 Ghz (Python)
75 a 38.76 % 55.08 % 31.81 % 0.4 s 1 core @ 2.5 Ghz (Python + C/C++)
76 Pseudo-LiDAR
This method uses stereo information.
 code 37.17 % 55.40 % 31.37 % 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.
77 Stereo R-CNN
This method uses stereo information.
 code 34.05 % 49.23 % 28.39 % 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.
78 34.04 % 49.68 % 28.45 %
79 SA_3D 31.21 % 41.70 % 25.96 % 0.3 s GPU @ 2.5 Ghz (Python)
80 RT3DStereo
This method uses stereo information.
 24.10 % 28.50 % 20.32 % 0.08 s GPU @ 2.5 Ghz (C/C++)
81 RT3D
This method makes use of Velodyne laser scans.
 21.27 % 23.49 % 19.81 % 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.
82 StereoFENet
This method uses stereo information.
 20.37 % 29.93 % 16.59 % 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.
83 monocular 16.08 % 21.48 % 15.26 % 0.4 s GPU @ 2.5 Ghz (Python + C/C++)
84 MonoDIS 15.12 % 11.81 % 12.71 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
85 DLMB
This method makes use of Velodyne laser scans.
This is an online method (no batch processing).
 14.49 % 15.16 % 12.94 % 0.03 s 8 cores @ 3.5 Ghz (C/C++)
86 BirdNet
This method makes use of Velodyne laser scans.
 13.44 % 14.75 % 12.04 % 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.
87 Mono3D_PLiDAR code 13.44 % 17.12 % 12.38 % 0.1 s NVIDIA GeForce 1080 (pytorch)
X. Weng and K. Kitani: Monocular 3D Object Detection with Pseudo-LiDAR Point Cloud. arXiv:1903.09847 2019.
88 MonoGRNet code 12.90 % 11.29 % 11.34 % 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.
89 Licar
This method makes use of Velodyne laser scans.
 12.88 % 16.25 % 13.67 % 0.09 s GPU @ 2.0 Ghz (Python)
90 TopNet-HighRes
This method makes use of Velodyne laser scans.
 12.58 % 15.29 % 12.25 % 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.
91 MonoPSR 10.85 % 12.57 % 9.06 % 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.
92 ROI-10D 10.30 % 12.30 % 9.39 % 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.
93 DT3D 9.92 % 15.37 % 9.26 % 0,21s GPU @ 2.5 Ghz (Python)
94 SS3D 9.58 % 11.74 % 7.77 % 48 ms Tesla V100 (Python)
E. Jörgensen, C. Zach and F. Kahl: Monocular 3D Object Detection and Box Fitting Trained End-to-End Using Intersection-over-Union Loss. CoRR 2019.
95 mylsi-faster-rcnn 9.49 % 11.80 % 9.19 % 0.3 s 1 core @ 2.5 Ghz (Python)
96 CSoR
This method makes use of Velodyne laser scans.
 6.79 % 6.76 % 6.14 % 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.
97 mymask-rcnn 6.65 % 10.90 % 6.34 % 0.3 s 1 core @ 2.5 Ghz (Python)
98 A3DODWTDA (image) code 6.45 % 6.76 % 4.87 % 0.8 s GPU @ 3.0 Ghz (Python)
F. Gustafsson and E. Linder-Norén: Automotive 3D Object Detection Without Target Domain Annotations. 2018.
99 MonoFENet 6.36 % 9.31 % 5.61 % 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.
100 GS3D 6.29 % 7.69 % 6.16 % 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.
101 Shift R-CNN (mono) code 5.22 % 8.13 % 4.78 % 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.
102 RAR-Net 4.22 % 6.55 % 3.26 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
103 TopNet-UncEst
This method makes use of Velodyne laser scans.
 3.93 % 6.21 % 3.78 % 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.
104 MF3D 3.17 % 3.81 % 3.25 % 0.03 s GPU @ 2.5 Ghz (C/C++)
105 OFT-Net 2.50 % 3.28 % 2.27 % 0.5 s 8 cores @ 2.5 Ghz (Python + C/C++)
106 FQNet 2.42 % 3.48 % 1.96 % 0.5 s 1 core @ 2.5 Ghz (Python)
107 3D-SSMFCNN code 2.28 % 2.39 % 1.52 % 0.1 s GPU @ 1.5 Ghz (C/C++)
L. Novak: Vehicle Detection and Pose Estimation for Autonomous Driving. 2017.
108 3DVSSD 1.14 % 1.38 % 1.27 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
109 monoref3d 0.55 % 0.53 % 0.55 % 0.1 s 1 core @ 2.5 Ghz (Python)
110 ref3D 0.00 % 0.00 % 0.00 % 0.1 s 1 core @ 2.5 Ghz (Python)
111 ref3D 0.00 % 0.00 % 0.00 % 0.1 s 1 core @ 2.5 Ghz (Python + C/C++)
112 mBoW
This method makes use of Velodyne laser scans.
 0.00 % 0.00 % 0.00 % 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.
Table as LaTeX | Only published Methods

Pedestrian


Method Setting Code Moderate Easy Hard Runtime Environment
1 A-VoxelNet 46.64 % 54.83 % 42.39 % 0.029 s GPU @ 2.5 Ghz (Python)
2 F-ConvNet
This method makes use of Velodyne laser scans.
 45.61 % 52.37 % 41.49 % 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. IROS 2019.
3 VMVS
This method makes use of Velodyne laser scans.
 45.01 % 53.98 % 41.72 % 0.25 s GPU @ 2.5 Ghz (Python)
J. Ku, A. Pon, S. Walsh and S. Waslander: Improving 3D object detection for pedestrians with virtual multi-view synthesis orientation estimation. IROS 2019.
4 F-PointNet
This method makes use of Velodyne laser scans.
 code 44.89 % 51.21 % 40.23 % 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.
5 IPOD 44.68 % 56.92 % 42.39 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
6 STD 44.24 % 53.08 % 41.97 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
7 epBRM
This method makes use of Velodyne laser scans.
 43.90 % 50.38 % 40.91 % 0.10 s 1 core @ 2.5 Ghz (C/C++)
8 PointPillars
This method makes use of Velodyne laser scans.
 code 43.53 % 52.08 % 41.49 % 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.
9 Multi-3D
This method makes use of Velodyne laser scans.
 42.87 % 51.17 % 38.94 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
10 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 42.81 % 50.80 % 40.88 % 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.
11 SECOND code 42.56 % 51.07 % 37.29 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
12 MDC
This method makes use of Velodyne laser scans.
 42.54 % 50.79 % 36.56 % 0.17 s GPU @ 2.5 Ghz (Python)
13 MMLab-PointRCNN
This method makes use of Velodyne laser scans.
 code 41.78 % 49.43 % 38.63 % 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. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2019.
14 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 41.64 % 49.83 % 39.28 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
15 ARPNET 41.62 % 50.00 % 39.19 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
16 SCANet 41.44 % 50.66 % 36.60 % 0.17 s >8 cores @ 2.5 Ghz (Python)
17 FOFNet
This method makes use of Velodyne laser scans.
 41.21 % 49.44 % 36.42 % 0.04 s GPU @ 2.5 Ghz (Python)
18 CONV-BOX
This method makes use of Velodyne laser scans.
 41.01 % 47.74 % 35.98 % 0.2 s Tesla V100
19 PP_v1.0 code 38.86 % 46.28 % 36.25 % 0.02s 1 core @ 2.5 Ghz (C/C++)
20 ELLIOT
This method makes use of Velodyne laser scans.
 37.78 % 45.94 % 34.94 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
21 CFR
This method makes use of Velodyne laser scans.
 36.86 % 44.64 % 35.57 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
22 anm 34.71 % 45.89 % 32.43 % 3 s 1 core @ 2.5 Ghz (C/C++)
23 anonymous
This method makes use of Velodyne laser scans.
 33.37 % 40.19 % 27.90 % 0.75 s GPU @ 3.5 Ghz (C/C++)
24 X_MD 33.23 % 40.34 % 28.19 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
25 SA_3D 32.58 % 39.20 % 27.65 % 0.3 s GPU @ 2.5 Ghz (Python)
26 CLF3D
This method makes use of Velodyne laser scans.
 31.65 % 35.85 % 26.94 % 0.13 s GPU @ 2.5 Ghz (Python)
27 AVOD
This method makes use of Velodyne laser scans.
 code 31.51 % 38.28 % 26.98 % 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.
28 31.30 % 38.00 % 28.77 %
29 a 21.85 % 28.14 % 20.92 % 0.4 s 1 core @ 2.5 Ghz (Python + C/C++)
30 Complexer-YOLO
This method makes use of Velodyne laser scans.
 15.32 % 19.45 % 14.80 % 0.06 s GPU @ 3.5 Ghz (C/C++)
M. Simon, K. Amende, A. Kraus, J. Honer, T. Samann, H. Kaulbersch, S. Milz and H. Michael Gross: Complexer-YOLO: Real-Time 3D Object Detection and Tracking on Semantic Point Clouds. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops 2019.
31 BirdNet
This method makes use of Velodyne laser scans.
 11.80 % 14.31 % 10.55 % 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 MonoPSR 10.66 % 12.65 % 10.08 % 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 Shift R-CNN (mono) code 10.59 % 13.36 % 10.59 % 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.
34 TopNet-HighRes
This method makes use of Velodyne laser scans.
 9.66 % 13.45 % 9.64 % 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.
35 TopNet-UncEst
This method makes use of Velodyne laser scans.
 4.55 % 6.19 % 4.55 % 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.
36 RT3DStereo
This method uses stereo information.
 4.25 % 4.27 % 4.26 % 0.08 s GPU @ 2.5 Ghz (C/C++)
37 SS3D 3.28 % 3.52 % 2.37 % 48 ms Tesla V100 (Python)
E. Jörgensen, C. Zach and F. Kahl: Monocular 3D Object Detection and Box Fitting Trained End-to-End Using Intersection-over-Union Loss. CoRR 2019.
38 mylsi-faster-rcnn 1.85 % 2.22 % 1.77 % 0.3 s 1 core @ 2.5 Ghz (Python)
39 mymask-rcnn 1.44 % 1.86 % 1.38 % 0.3 s 1 core @ 2.5 Ghz (Python)
40 DT3D 1.14 % 1.14 % 1.14 % 0,21s GPU @ 2.5 Ghz (Python)
41 OFT-Net 1.11 % 1.06 % 1.06 % 0.5 s 8 cores @ 2.5 Ghz (Python + C/C++)
42 mBoW
This method makes use of Velodyne laser scans.
 0.00 % 0.00 % 0.00 % 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.
Table as LaTeX | Only published Methods

Cyclist


Method Setting Code Moderate Easy Hard Runtime Environment
1 F-ConvNet
This method makes use of Velodyne laser scans.
 64.68 % 79.58 % 57.03 % 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. IROS 2019.
2 MMLab-PartA^2
This method makes use of Velodyne laser scans.
 62.73 % 78.58 % 57.74 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
S. Shi, Z. Wang, X. Wang and H. Li: Part-A^2 Net: 3D Part-Aware and Aggregation Neural Network for Object Detection from Point Cloud. arXiv preprint arXiv:1907.03670 2019.
3 STD 62.53 % 78.89 % 55.77 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
4 FOFNet
This method makes use of Velodyne laser scans.
 59.65 % 75.36 % 53.03 % 0.04 s GPU @ 2.5 Ghz (Python)
5 MMLab-PointRCNN
This method makes use of Velodyne laser scans.
 code 59.60 % 73.93 % 53.59 % 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. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2019.
6 Multi-3D
This method makes use of Velodyne laser scans.
 59.40 % 75.99 % 51.50 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
7 ARPNET 59.12 % 72.29 % 53.35 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
8 PointPillars
This method makes use of Velodyne laser scans.
 code 59.07 % 75.78 % 52.92 % 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.
9 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 58.19 % 72.12 % 51.54 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
10 MDC
This method makes use of Velodyne laser scans.
 57.27 % 75.27 % 49.75 % 0.17 s GPU @ 2.5 Ghz (Python)
11 epBRM
This method makes use of Velodyne laser scans.
 56.94 % 70.52 % 51.70 % 0.10 s 1 core @ 2.5 Ghz (C/C++)
12 A-VoxelNet 56.86 % 70.65 % 50.76 % 0.029 s GPU @ 2.5 Ghz (Python)
13 F-PointNet
This method makes use of Velodyne laser scans.
 code 56.77 % 71.96 % 50.39 % 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.
14 CONV-BOX
This method makes use of Velodyne laser scans.
 54.45 % 68.27 % 52.26 % 0.2 s Tesla V100
15 SECOND code 53.85 % 70.51 % 46.90 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
16 IPOD 53.46 % 71.40 % 48.34 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
17 SCANet 53.07 % 67.97 % 50.81 % 0.17 s >8 cores @ 2.5 Ghz (Python)
18 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 52.18 % 64.00 % 46.61 % 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.
19 ELLIOT
This method makes use of Velodyne laser scans.
 51.17 % 68.87 % 46.35 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
20 CFR
This method makes use of Velodyne laser scans.
 50.73 % 65.70 % 44.93 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
21 PP_v1.0 code 50.60 % 66.86 % 44.84 % 0.02s 1 core @ 2.5 Ghz (C/C++)
22 AVOD
This method makes use of Velodyne laser scans.
 code 44.90 % 60.11 % 38.80 % 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.
23 X_MD 37.22 % 51.69 % 36.44 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
24 anm 35.86 % 50.06 % 31.11 % 3 s 1 core @ 2.5 Ghz (C/C++)
25 CLF3D
This method makes use of Velodyne laser scans.
 35.39 % 50.58 % 33.55 % 0.13 s GPU @ 2.5 Ghz (Python)
26 Complexer-YOLO
This method makes use of Velodyne laser scans.
 23.48 % 28.36 % 22.85 % 0.06 s GPU @ 3.5 Ghz (C/C++)
M. Simon, K. Amende, A. Kraus, J. Honer, T. Samann, H. Kaulbersch, S. Milz and H. Michael Gross: Complexer-YOLO: Real-Time 3D Object Detection and Tracking on Semantic Point Clouds. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops 2019.
27 a 21.25 % 32.66 % 19.77 % 0.4 s 1 core @ 2.5 Ghz (Python + C/C++)
28 SA_3D 14.15 % 17.99 % 13.52 % 0.3 s GPU @ 2.5 Ghz (Python)
29 BirdNet
This method makes use of Velodyne laser scans.
 12.43 % 18.35 % 11.88 % 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.
30 MonoPSR 11.01 % 13.43 % 9.93 % 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.
31 SS3D 9.09 % 10.84 % 9.09 % 48 ms Tesla V100 (Python)
E. Jörgensen, C. Zach and F. Kahl: Monocular 3D Object Detection and Box Fitting Trained End-to-End Using Intersection-over-Union Loss. CoRR 2019.
32 TopNet-UncEst
This method makes use of Velodyne laser scans.
 7.36 % 8.53 % 6.93 % 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.
33 RT3DStereo
This method uses stereo information.
 6.63 % 6.62 % 4.03 % 0.08 s GPU @ 2.5 Ghz (C/C++)
34 TopNet-HighRes
This method makes use of Velodyne laser scans.
 5.98 % 4.48 % 6.18 % 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.
35 Shift R-CNN (mono) code 3.03 % 3.03 % 3.03 % 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.
36 mylsi-faster-rcnn 1.68 % 2.53 % 1.41 % 0.3 s 1 core @ 2.5 Ghz (Python)
37 DT3D 1.20 % 1.76 % 1.26 % 0,21s GPU @ 2.5 Ghz (Python)
38 mymask-rcnn 0.84 % 1.18 % 0.83 % 0.3 s 1 core @ 2.5 Ghz (Python)
39 OFT-Net 0.43 % 0.43 % 0.43 % 0.5 s 8 cores @ 2.5 Ghz (Python + C/C++)
40 mBoW
This method makes use of Velodyne laser scans.
 0.00 % 0.00 % 0.00 % 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.
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