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

Bird's Eye View Evaluation 2017


The bird's eye view 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 bird's eye view 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 a bounding box overlap of 70% in bird's eye view, while for pedestrians and cyclists we require an 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 Fusion-RCNN 88.35 % 89.59 % 79.56 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
2 STD 87.76 % 89.66 % 86.89 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
3 Alibaba-AILabsX
This method makes use of Velodyne laser scans.
 87.49 % 89.10 % 78.98 % 0.05 s 1 core @ >3.5 Ghz (C/C++)
4 UberATG-MMF
This method makes use of Velodyne laser scans.
 87.47 % 89.49 % 79.10 % 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.
5 Patches - EMP
This method makes use of Velodyne laser scans.
 87.45 % 89.96 % 83.63 % 0.5 s GPU @ 2.5 Ghz (Python)
6 GPOD
This method makes use of Velodyne laser scans.
 87.39 % 89.54 % 81.65 % 0.1 s GPU @ 2.5 Ghz (Python)
7 Part-A^2
This method makes use of Velodyne laser scans.
 87.04 % 89.74 % 79.52 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
8 A-VoxelNet 86.96 % 88.30 % 78.44 % 0.029 s GPU @ 2.5 Ghz (Python)
9 MPNet
This method makes use of Velodyne laser scans.
 86.95 % 89.28 % 80.32 % 0.02 s GPU @ 2.5 Ghz (Python + C/C++)
10 MMV 86.90 % 88.61 % 78.45 % 0.4 s GPU @ 2.5 Ghz (C/C++)
11 MODet
This method makes use of Velodyne laser scans.
 86.80 % 89.15 % 79.62 % 0.05 s GTX1080Ti
12 SRF 86.75 % 88.60 % 78.89 % 0.05 s GPU @ 2.5 Ghz (Python + C/C++)
13 Alibaba-AILabsX
This method makes use of Velodyne laser scans.
 86.71 % 88.84 % 78.62 % 0.2 s GPU @ >3.5 Ghz (Python)
14 ARPNET 86.67 % 88.10 % 77.94 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
15 SECOND-V1.5
This method makes use of Velodyne laser scans.
 code 86.60 % 88.72 % 78.51 % 0.04 s GPU @ 2.0 Ghz (Python + C/C++)
16 UberATG-HDNET
This method makes use of Velodyne laser scans.
 86.57 % 89.14 % 78.32 % 0.05 s GPU @ 2.5 Ghz (Python)
B. Yang, M. Liang and R. Urtasun: HDNET: Exploiting HD Maps for 3D Object Detection. 2nd Conference on Robot Learning (CoRL) 2018.
17 Patches
This method makes use of Velodyne laser scans.
 86.55 % 89.78 % 79.22 % 0.15 s GPU @ 2.0 Ghz
18 RAL
This method makes use of Velodyne laser scans.
 86.39 % 88.16 % 78.66 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
19 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 86.29 % 88.67 % 79.27 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
20 TBA 86.29 % 88.75 % 80.13 % 0.07 s 1 core @ 2.5 Ghz (Python)
21 NU-optim 86.21 % 88.39 % 78.34 % 0.04 s GPU @ >3.5 Ghz (Python)
22 SegVoxelNet 86.16 % 88.62 % 78.68 % 0.04 s 1 core @ 2.5 Ghz (Python)
23 PC-CNN-V2
This method makes use of Velodyne laser scans.
 86.10 % 88.49 % 77.26 % 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.
24 PointPillars
This method makes use of Velodyne laser scans.
 code 86.10 % 88.35 % 79.83 % 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.
25 Fast Point R-CNNv1.1
This method makes use of Velodyne laser scans.
 86.10 % 88.03 % 78.17 % 0.06 s GPU @ 2.5 Ghz (Python + C/C++)
26 DH-ARI 86.08 % 87.93 % 78.22 % 0.2 s 1 core @ >3.5 Ghz (Python + C/C++)
27 epBRM
This method makes use of Velodyne laser scans.
 86.08 % 88.75 % 78.80 % 0.1 s GPU @ >3.5 Ghz (Python + C/C++)
28 PointRCNN-deprecated
This method makes use of Velodyne laser scans.
 86.04 % 89.28 % 79.02 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
29 MVX-Net
This method makes use of Velodyne laser scans.
 85.89 % 89.15 % 78.07 % 0.06 s GPU @ 3.0 Ghz (Python + C/C++)
30 UberATG-ContFuse
This method makes use of Velodyne laser scans.
 85.83 % 88.81 % 77.33 % 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.
31 MDC
This method makes use of Velodyne laser scans.
 85.68 % 88.65 % 77.03 % 0.17 s GPU @ 2.5 Ghz (Python)
32 PointRCNN
This method makes use of Velodyne laser scans.
 code 85.68 % 89.47 % 79.10 % 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.
33 Fast Point R-CNN
This method makes use of Velodyne laser scans.
 85.53 % 87.98 % 77.68 % 0.06 s GPU @ 2.5 Ghz (Python + C/C++)
34 AVOD
This method makes use of Velodyne laser scans.
 code 85.44 % 86.80 % 77.73 % 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.
35 SeoulRobotics-HFD
This method makes use of Velodyne laser scans.
 85.22 % 87.17 % 77.71 % 0.035 s GPU (C++)
36 Roadstar.ai 85.04 % 89.95 % 78.75 % 0.08 s GPU @ 2.0 Ghz (Python)
37 CONV-BOX
This method makes use of Velodyne laser scans.
 84.56 % 87.54 % 77.79 % 0.2 s Tesla V100
38 F-PointNet
This method makes use of Velodyne laser scans.
 code 84.00 % 88.70 % 75.33 % 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.
39 IPOD 83.98 % 86.93 % 77.85 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
40 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 83.79 % 88.53 % 77.90 % 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.
41 UberATG-PIXOR++
This method makes use of Velodyne laser scans.
 83.70 % 89.38 % 77.97 % 0.035 s GPU @ 2.5 Ghz (Python)
B. Yang, M. Liang and R. Urtasun: HDNET: Exploiting HD Maps for 3D Object Detection. 2nd Conference on Robot Learning (CoRL) 2018.
42 F-ConvNet
This method makes use of Velodyne laser scans.
 83.08 % 89.69 % 74.56 % 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.
43 PP_v1.0 code 81.12 % 87.72 % 78.13 % 0.02s 1 core @ 2.5 Ghz (C/C++)
44 PAD 80.51 % 87.87 % 78.36 % 0.15 s 1 core @ 2.5 Ghz (Python)
45 Voxel-FPN 80.17 % 88.96 % 79.14 % 0.02 s GPU @ 2.5 Ghz (Python + C/C++)
46 PFPN 80.14 % 89.05 % 79.02 % 0.02 s 4 cores @ >3.5 Ghz (Python)
47 yl_net 80.06 % 85.64 % 77.52 % 0.03 s GPU @ 2.5 Ghz (Python)
48 tiny_rfdet code 80.05 % 86.53 % 76.28 % 0.01 s GPU @ 2.5 Ghz (Python)
49 MVSLN 79.71 % 89.08 % 78.26 % 0.1s s 1 core @ 2.5 Ghz (C/C++)
50 CFR
This method makes use of Velodyne laser scans.
 79.68 % 88.77 % 78.21 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
51 3DBN
This method makes use of Velodyne laser scans.
 79.40 % 88.13 % 77.97 % 0.13s 1080Ti (Python+C/C++)
X. Li, J. Guivant, N. Kwok and Y. Xu: 3D Backbone Network for 3D Object Detection. CoRR 2019.
52 SECOND code 79.37 % 88.07 % 77.95 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
53 ELLIOT
This method makes use of Velodyne laser scans.
 79.30 % 87.06 % 77.28 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
54 AILabs3D
This method makes use of Velodyne laser scans.
 79.09 % 88.32 % 77.40 % 0.6 s GPU @ >3.5 Ghz (Python)
55 SCANet 79.01 % 87.02 % 77.20 % 0.17 s >8 cores @ 2.5 Ghz (Python)
56 DFD 78.92 % 86.98 % 77.22 % 0.05 s GPU @ 2.0 Ghz (Python + C/C++)
57 SCANet 78.64 % 87.36 % 77.37 % 0.09s GPU @ 2.5 Ghz (Python)
58 SECA 78.54 % 87.52 % 77.25 % 0.09 s GPU @ 2.5 Ghz (Python)
59 FNV1_RPN 78.51 % 87.40 % 70.23 % 0.12 s 1 core @ 2.5 Ghz (Python + C/C++)
60 SECA 78.50 % 87.65 % 76.99 % 1 s GPU @ 2.5 Ghz (Python)
61 VSE 78.50 % 87.65 % 76.99 % 0.15 s GPU @ 2.5 Ghz (Python)
62 NLK 78.28 % 87.65 % 76.20 % 0.02 s 1 core @ 2.5 Ghz (Python)
63 FNV1_Fusion 78.23 % 87.07 % 76.72 % 0.11 s GPU @ 2.5 Ghz (Python)
64 VoxelNet(Unofficial) 77.94 % 86.60 % 69.66 % 0.5 s GPU @ 2.0 Ghz (Python)
65 X_MD 77.80 % 87.02 % 69.66 % 0.2 s 1 core @ 2.5 Ghz (Python + C/C++)
66 RTL3D 77.75 % 86.95 % 76.08 % 0.02 s GPU @ 2.5 Ghz (Python)
67 FNV1 77.69 % 85.53 % 69.46 % 0.11 s GPU @ 2.5 Ghz (Python)
68 UberATG-PIXOR
This method makes use of Velodyne laser scans.
 77.05 % 81.70 % 72.95 % 0.035 s TITAN Xp (Python)
B. Yang, W. Luo and R. Urtasun: PIXOR: Real-time 3D Object Detection from Point Clouds. CVPR 2018.
69 MV3D (LIDAR)
This method makes use of Velodyne laser scans.
 77.00 % 85.82 % 68.94 % 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.
70 MV3D
This method makes use of Velodyne laser scans.
 76.90 % 86.02 % 68.49 % 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.
71 Multi-3D
This method makes use of Velodyne laser scans.
 76.65 % 85.72 % 66.79 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
72 FNV2 74.36 % 80.68 % 65.88 % 0.18 s GPU @ 2.5 Ghz (Python)
73 LaserNet 73.77 % 78.25 % 66.47 % 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.
74 CLF3D
This method makes use of Velodyne laser scans.
 73.01 % 80.09 % 58.56 % 0.13 s GPU @ 2.5 Ghz (Python)
75 A3DODWTDA
This method makes use of Velodyne laser scans.
 code 72.86 % 76.65 % 64.51 % 0.08 s GPU @ 3.0 Ghz (Python)
F. Gustafsson and E. Linder-Norén: Automotive 3D Object Detection Without Target Domain Annotations. 2018.
76 anm 72.73 % 81.79 % 63.98 % 3 s 1 core @ 2.5 Ghz (C/C++)
77 avodC 69.53 % 84.54 % 67.98 % 0.1 s GPU @ 2.5 Ghz (Python)
78 TopNet-Retina
This method makes use of Velodyne laser scans.
 67.14 % 76.62 % 65.92 % 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.
79 Complexer-YOLO
This method makes use of Velodyne laser scans.
 66.07 % 74.23 % 65.70 % 0.06 s GPU @ 3.5 Ghz (C/C++)
80 TopNet-DecayRate
This method makes use of Velodyne laser scans.
 64.12 % 79.76 % 56.48 % 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.
81 3D FCN
This method makes use of Velodyne laser scans.
 62.54 % 69.94 % 55.94 % >5 s 1 core @ 2.5 Ghz (C/C++)
B. Li: 3D Fully Convolutional Network for Vehicle Detection in Point Cloud. IROS 2017.
82 TopNet-UncEst
This method makes use of Velodyne laser scans.
 62.42 % 70.13 % 55.20 % 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.
83 Pseudo-LiDAR V2
This method uses stereo information.
 57.49 % 75.97 % 53.61 % 0.4 s GPU @ 2.5 Ghz (Python)
84 TopNet-HighRes
This method makes use of Velodyne laser scans.
 53.71 % 67.53 % 46.54 % 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.
85 BirdNet
This method makes use of Velodyne laser scans.
 50.81 % 75.52 % 50.00 % 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.
86 RT3DStereo
This method uses stereo information.
 49.48 % 59.32 % 43.16 % 0.08 s GPU @ 2.5 Ghz (C/C++)
87 Pseudo-LiDAR
This method uses stereo information.
 code 47.20 % 66.83 % 40.30 % 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.
88 SA_3D 47.00 % 58.98 % 39.47 % 0.3 s GPU @ 2.5 Ghz (Python)
89 45.11 % 65.08 % 38.42 %
90 Stereo R-CNN
This method uses stereo information.
 code 43.87 % 61.67 % 36.44 % 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.
91 RT3D
This method makes use of Velodyne laser scans.
 42.10 % 54.68 % 44.05 % 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.
92 Licar
This method makes use of Velodyne laser scans.
 40.40 % 45.81 % 37.09 % 0.09 s GPU @ 2.0 Ghz (Python)
93 DLMB
This method makes use of Velodyne laser scans.
This is an online method (no batch processing).
 37.10 % 45.12 % 32.88 % 0.03 s 8 cores @ 3.5 Ghz (C/C++)
94 StereoFENet
This method uses stereo information.
 34.62 % 49.14 % 28.41 % 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.
95 monocular 22.24 % 27.91 % 18.62 % 0.4 s GPU @ 2.5 Ghz (Python + C/C++)
96 MonoDIS 19.08 % 18.88 % 17.41 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
97 CSoR
This method makes use of Velodyne laser scans.
 18.69 % 23.94 % 16.30 % 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.
98 MonoPSR 17.66 % 20.25 % 15.78 % 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.
99 DT3D 17.19 % 23.38 % 13.86 % 0,21s GPU @ 2.5 Ghz (Python)
100 MonoGRNet code 16.37 % 20.55 % 15.16 % 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.
101 mylsi-faster-rcnn 15.99 % 21.09 % 14.56 % 0.3 s 1 core @ 2.5 Ghz (Python)
102 SS3D 12.81 % 17.86 % 12.28 % 48 ms Tesla V100 (Python)
103 MonoFENet 12.71 % 18.08 % 10.55 % 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.
104 ROI-10D 12.40 % 16.77 % 11.39 % 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.
105 mymask-rcnn 12.17 % 19.13 % 10.56 % 0.3 s 1 core @ 2.5 Ghz (Python)
106 A3DODWTDA (image) code 10.21 % 10.61 % 8.64 % 0.8 s GPU @ 3.0 Ghz (Python)
F. Gustafsson and E. Linder-Norén: Automotive 3D Object Detection Without Target Domain Annotations. 2018.
107 GS3D 9.12 % 11.30 % 7.23 % 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.
108 Shift R-CNN 8.49 % 13.32 % 6.40 % 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.
109 OFT-Net 7.99 % 9.50 % 7.51 % 0.5 s 8 cores @ 2.5 Ghz (Python + C/C++)
110 RAR-Net 6.21 % 10.38 % 5.43 % 0.5 s 1 core @ 2.5 Ghz (C/C++)
111 MF3D 5.57 % 7.88 % 5.08 % 0.03 s GPU @ 2.5 Ghz (C/C++)
112 FQNet 4.62 % 6.51 % 3.99 % 0.5 s 1 core @ 2.5 Ghz (Python)
113 monoref3d 3.46 % 4.74 % 2.68 % 0.1 s 1 core @ 2.5 Ghz (Python)
114 ref3D 3.46 % 4.74 % 2.68 % 0.1 s 1 core @ 2.5 Ghz (Python + C/C++)
115 3D-SSMFCNN code 3.19 % 3.66 % 3.45 % 0.1 s GPU @ 1.5 Ghz (C/C++)
L. Novak: Vehicle Detection and Pose Estimation for Autonomous Driving. 2017.
116 3DVSSD 2.01 % 2.02 % 1.68 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
117 VeloFCN
This method makes use of Velodyne laser scans.
 0.33 % 0.15 % 0.47 % 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 .
118 ref3D 0.00 % 0.00 % 0.01 % 0.1 s 1 core @ 2.5 Ghz (Python)
119 multi-task CNN 0.00 % 0.00 % 0.00 % 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.
120 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 51.90 % 60.44 % 49.48 % 0.029 s GPU @ 2.5 Ghz (Python)
2 VMVS
This method makes use of Velodyne laser scans.
 51.73 % 61.46 % 47.69 % 0.25 s GPU @ 2.5 Ghz (Python)
3 STD 51.39 % 60.99 % 45.89 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
4 IPOD 51.24 % 60.83 % 45.40 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
5 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 51.05 % 58.75 % 47.54 % 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.
6 F-ConvNet
This method makes use of Velodyne laser scans.
 50.48 % 58.90 % 46.72 % 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.
7 PointPillars
This method makes use of Velodyne laser scans.
 code 50.23 % 58.66 % 47.19 % 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.
8 F-PointNet
This method makes use of Velodyne laser scans.
 code 50.22 % 58.09 % 47.20 % 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.
9 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 48.93 % 56.89 % 46.28 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
10 Multi-3D
This method makes use of Velodyne laser scans.
 47.78 % 54.91 % 43.34 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
11 PointRCNN
This method makes use of Velodyne laser scans.
 code 47.53 % 55.92 % 44.67 % 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.
12 epBRM
This method makes use of Velodyne laser scans.
 47.17 % 53.24 % 44.83 % 0.10 s 1 core @ 2.5 Ghz (C/C++)
13 PP_v1.0 code 46.48 % 54.26 % 44.14 % 0.02s 1 core @ 2.5 Ghz (C/C++)
14 GPOD
This method makes use of Velodyne laser scans.
 46.39 % 53.09 % 43.62 % 0.1 s GPU @ 2.5 Ghz (Python)
15 SECOND code 46.27 % 55.10 % 44.76 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
16 Roadstar.ai 46.22 % 49.94 % 44.14 % 0.08 s GPU @ 2.0 Ghz (Python)
17 CONV-BOX
This method makes use of Velodyne laser scans.
 45.09 % 52.71 % 43.90 % 0.2 s Tesla V100
18 MDC
This method makes use of Velodyne laser scans.
 45.02 % 53.51 % 43.67 % 0.17 s GPU @ 2.5 Ghz (Python)
19 SCANet 44.30 % 53.68 % 42.65 % 0.17 s >8 cores @ 2.5 Ghz (Python)
20 ARPNET 43.35 % 50.80 % 37.79 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
21 ELLIOT
This method makes use of Velodyne laser scans.
 42.66 % 50.68 % 39.67 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
22 anm 40.48 % 50.07 % 35.64 % 3 s 1 core @ 2.5 Ghz (C/C++)
23 CFR
This method makes use of Velodyne laser scans.
 39.24 % 52.47 % 38.07 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
24 SA_3D 35.93 % 47.45 % 34.99 % 0.3 s GPU @ 2.5 Ghz (Python)
25 anonymous
This method makes use of Velodyne laser scans.
 35.58 % 42.50 % 35.06 % 0.75 s GPU @ 3.5 Ghz (C/C++)
26 X_MD 35.57 % 43.28 % 34.25 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
27 35.26 % 43.26 % 32.85 %
28 AVOD
This method makes use of Velodyne laser scans.
 code 35.24 % 42.51 % 33.97 % 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 CLF3D
This method makes use of Velodyne laser scans.
 34.25 % 43.11 % 33.01 % 0.13 s GPU @ 2.5 Ghz (Python)
30 BirdNet
This method makes use of Velodyne laser scans.
 21.35 % 26.07 % 19.96 % 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.
31 Complexer-YOLO
This method makes use of Velodyne laser scans.
 20.88 % 22.00 % 20.81 % 0.06 s GPU @ 3.5 Ghz (C/C++)
32 TopNet-HighRes
This method makes use of Velodyne laser scans.
 19.08 % 24.30 % 18.46 % 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.
33 TopNet-Retina
This method makes use of Velodyne laser scans.
 17.62 % 18.28 % 14.08 % 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.
34 TopNet-DecayRate
This method makes use of Velodyne laser scans.
 12.59 % 15.09 % 12.23 % 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.
35 TopNet-UncEst
This method makes use of Velodyne laser scans.
 11.63 % 12.39 % 11.39 % 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 Shift R-CNN 11.44 % 13.81 % 10.76 % 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.
37 MonoPSR 11.22 % 14.27 % 10.54 % 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.
38 RT3DStereo
This method uses stereo information.
 5.30 % 5.39 % 5.19 % 0.08 s GPU @ 2.5 Ghz (C/C++)
39 SS3D 3.52 % 3.86 % 2.50 % 48 ms Tesla V100 (Python)
40 mylsi-faster-rcnn 3.52 % 4.56 % 3.52 % 0.3 s 1 core @ 2.5 Ghz (Python)
41 mymask-rcnn 2.64 % 3.91 % 2.63 % 0.3 s 1 core @ 2.5 Ghz (Python)
42 OFT-Net 1.55 % 1.93 % 1.65 % 0.5 s 8 cores @ 2.5 Ghz (Python + C/C++)
43 DT3D 1.22 % 1.15 % 1.14 % 0,21s GPU @ 2.5 Ghz (Python)
44 mBoW
This method makes use of Velodyne laser scans.
 0.01 % 0.01 % 0.01 % 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.
 68.62 % 82.59 % 60.62 % 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 66.77 % 81.52 % 60.78 % 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 STD 65.32 % 81.04 % 57.85 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
4 MDC
This method makes use of Velodyne laser scans.
 64.78 % 77.38 % 56.77 % 0.17 s GPU @ 2.5 Ghz (Python)
5 Multi-3D
This method makes use of Velodyne laser scans.
 63.72 % 78.76 % 55.50 % 0.15 s 1 core @ 2.5 Ghz (C/C++)
6 ARPNET 62.76 % 77.99 % 55.55 % 0.08 s GPU @ 2.5 Ghz (Python + C/C++)
7 Tencent_ADlab_Lidar
This method makes use of Velodyne laser scans.
 62.33 % 78.15 % 55.87 % 0.1 s GPU @ 2.5 Ghz (Python + C/C++)
8 PointPillars
This method makes use of Velodyne laser scans.
 code 62.25 % 79.14 % 56.00 % 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 F-PointNet
This method makes use of Velodyne laser scans.
 code 61.96 % 75.38 % 54.68 % 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.
10 CONV-BOX
This method makes use of Velodyne laser scans.
 61.84 % 71.60 % 55.03 % 0.2 s Tesla V100
11 A-VoxelNet 60.85 % 76.71 % 54.27 % 0.029 s GPU @ 2.5 Ghz (Python)
12 GPOD
This method makes use of Velodyne laser scans.
 60.73 % 68.84 % 55.05 % 0.1 s GPU @ 2.5 Ghz (Python)
13 Roadstar.ai 60.26 % 67.31 % 54.62 % 0.08 s GPU @ 2.0 Ghz (Python)
14 epBRM
This method makes use of Velodyne laser scans.
 60.22 % 73.33 % 54.10 % 0.10 s 1 core @ 2.5 Ghz (C/C++)
15 SCANet 59.09 % 70.73 % 52.79 % 0.17 s >8 cores @ 2.5 Ghz (Python)
16 IPOD 58.92 % 77.10 % 51.01 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
17 ELLIOT
This method makes use of Velodyne laser scans.
 57.63 % 76.27 % 52.36 % 0.1 s 1 core @ 2.5 Ghz (C/C++)
18 AVOD-FPN
This method makes use of Velodyne laser scans.
 code 57.48 % 68.09 % 50.77 % 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.
19 SECOND code 56.04 % 73.67 % 48.78 % 38 ms 1080Ti
Y. Yan, Y. Mao and B. Li: SECOND: Sparsely Embedded Convolutional Detection. Sensors 2018.
20 PP_v1.0 code 54.59 % 70.83 % 49.11 % 0.02s 1 core @ 2.5 Ghz (C/C++)
21 CFR
This method makes use of Velodyne laser scans.
 53.41 % 69.05 % 46.81 % 0.06 s 1 core @ 2.5 Ghz (C/C++)
22 AVOD
This method makes use of Velodyne laser scans.
 code 47.74 % 63.66 % 46.55 % 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 44.49 % 54.22 % 38.48 % 0.2 s 1 core @ 2.5 Ghz (C/C++)
24 CLF3D
This method makes use of Velodyne laser scans.
 41.57 % 53.29 % 35.80 % 0.13 s GPU @ 2.5 Ghz (Python)
25 anm 40.42 % 56.84 % 35.20 % 3 s 1 core @ 2.5 Ghz (C/C++)
26 TopNet-Retina
This method makes use of Velodyne laser scans.
 38.11 % 50.15 % 36.49 % 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.
27 Complexer-YOLO
This method makes use of Velodyne laser scans.
 30.16 % 36.12 % 26.01 % 0.06 s GPU @ 3.5 Ghz (C/C++)
28 BirdNet
This method makes use of Velodyne laser scans.
 27.18 % 38.93 % 25.51 % 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.
29 SA_3D 20.56 % 26.45 % 16.42 % 0.3 s GPU @ 2.5 Ghz (Python)
30 TopNet-DecayRate
This method makes use of Velodyne laser scans.
 19.92 % 28.06 % 19.13 % 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.
31 TopNet-HighRes
This method makes use of Velodyne laser scans.
 12.45 % 15.70 % 12.76 % 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.
32 MonoPSR 12.17 % 14.75 % 11.35 % 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 TopNet-UncEst
This method makes use of Velodyne laser scans.
 10.34 % 14.80 % 10.42 % 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.
34 SS3D 9.65 % 11.52 % 9.09 % 48 ms Tesla V100 (Python)
35 RT3DStereo
This method uses stereo information.
 7.59 % 7.70 % 7.51 % 0.08 s GPU @ 2.5 Ghz (C/C++)
36 Shift R-CNN 3.03 % 3.58 % 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.
37 mylsi-faster-rcnn 2.75 % 3.06 % 2.57 % 0.3 s 1 core @ 2.5 Ghz (Python)
38 DT3D 1.26 % 2.54 % 1.47 % 0,21s GPU @ 2.5 Ghz (Python)
39 mymask-rcnn 1.12 % 2.20 % 1.18 % 0.3 s 1 core @ 2.5 Ghz (Python)
40 OFT-Net 0.43 % 0.79 % 0.43 % 0.5 s 8 cores @ 2.5 Ghz (Python + C/C++)
41 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