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

Optical Flow Evaluation 2015


The stereo 2015 / flow 2015 / scene flow 2015 benchmark consists of 200 training scenes and 200 test scenes (4 color images per scene, saved in loss less png format). Compared to the stereo 2012 and flow 2012 benchmarks, it comprises dynamic scenes for which the ground truth has been established in a semi-automatic process. Our evaluation server computes the percentage of bad pixels averaged over all ground truth pixels of all 200 test images. For this benchmark, we consider a pixel to be correctly estimated if the disparity or flow end-point error is <3px or <5% (for scene flow this criterion needs to be fulfilled for both disparity maps and the flow map). 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 disparity maps and flow fields. More details can be found in Object Scene Flow for Autonomous Vehicles (CVPR 2015).

Our evaluation table ranks all methods according to the number of erroneous pixels. All methods providing less than 100 % density have been interpolated using simple background interpolation as explained in the corresponding header file in the development kit. Legend:

  • D1: Percentage of stereo disparity outliers in first frame
  • D2: Percentage of stereo disparity outliers in second frame
  • Fl: Percentage of optical flow outliers
  • SF: Percentage of scene flow outliers (=outliers in either D0, D1 or Fl)
  • bg: Percentage of outliers averaged only over background regions
  • fg: Percentage of outliers averaged only over foreground regions
  • all: Percentage of outliers averaged over all ground truth pixels


Note: On 13.03.2017 we have fixed several small errors in the flow (noc+occ) ground truth of the dynamic foreground objects and manually verified all images for correctness by warping them according to the ground truth. As a consequence, all error numbers have decreased slightly. Please download the devkit and the annotations with the improved ground truth for the training set again if you have downloaded the files prior to 13.03.2017 and consider reporting these new number in all future publications. The last leaderboards before these corrections can be found here (optical flow 2015) and here (scene flow 2015). The leaderboards for the KITTI 2015 stereo benchmarks did not change.

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
  • Multiview: Method uses more than 2 temporally adjacent images
  • Motion stereo: Method uses epipolar geometry for computing optical flow
  • Additional training data: Use of additional data sources for training (see details)

Evaluation ground truth        Evaluation area

Method Setting Code Fl-bg Fl-fg Fl-all Density Runtime Environment
1 UberATG-DRISF
This method uses stereo information.
 3.59 % 10.40 % 4.73 % 100.00 % 0.75 s CPU+GPU @ 2.5 Ghz (Python)
W. Ma, S. Wang, R. Hu, Y. Xiong and R. Urtasun: Deep Rigid Instance Scene Flow. CVPR 2019.
2 DH-SF
This method uses stereo information.
 4.12 % 12.07 % 5.45 % 100.00 % 350 s 1 core @ 2.5 Ghz (Matlab + C/C++)
3 ISF
This method uses stereo information.
 5.40 % 10.29 % 6.22 % 100.00 % 10 min 1 core @ 3 Ghz (C/C++)
A. Behl, O. Jafari, S. Mustikovela, H. Alhaija, C. Rother and A. Geiger: Bounding Boxes, Segmentations and Object Coordinates: How Important is Recognition for 3D Scene Flow Estimation in Autonomous Driving Scenarios?. International Conference on Computer Vision (ICCV) 2017.
4 VCN 6.01 % 8.51 % 6.42 % 100.00 % 0.3 s Titan X Pascal
5 HD^3-Flow code 6.05 % 9.02 % 6.55 % 100.00 % 0.10 s NVIDIA Pascal Titan XP
Z. Yin, T. Darrell and F. Yu: Hierarchical Discrete Distribution Decomposition for Match Density Estimation. CVPR 2019.
6 PRSM
This method uses stereo information.
This method makes use of multiple (>2) views.
 code 5.33 % 13.40 % 6.68 % 100.00 % 300 s 1 core @ 2.5 Ghz (C/C++)
C. Vogel, K. Schindler and S. Roth: 3D Scene Flow Estimation with a Piecewise Rigid Scene Model. ijcv 2015.
7 OSF+TC
This method uses stereo information.
This method makes use of multiple (>2) views.
 5.76 % 13.31 % 7.02 % 100.00 % 50 min 1 core @ 2.5 Ghz (C/C++)
M. Neoral and J. Šochman: Object Scene Flow with Temporal Consistency. 22nd Computer Vision Winter Workshop (CVWW) 2017.
8 SSF
This method uses stereo information.
 5.63 % 14.71 % 7.14 % 100.00 % 5 min 1 core @ 2.5 Ghz (Matlab + C/C++)
Z. Ren, D. Sun, J. Kautz and E. Sudderth: Cascaded Scene Flow Prediction using Semantic Segmentation. International Conference on 3D Vision (3DV) 2017.
9 SPOSF
This method uses stereo information.
 5.41 % 15.96 % 7.16 % 100.00 % 10 min 1 core @ 3.5 Ghz (Matlab + C/C++)
10 MFF
This method makes use of multiple (>2) views.
 7.15 % 7.25 % 7.17 % 100.00 % 0.05 s NVIDIA Pascal Titan X (Python)
Z. Ren, O. Gallo, D. Sun, M. Yang, E. Sudderth and J. Kautz: A Fusion Approach for Multi-Frame Optical Flow Estimation. IEEE Winter Conference on Applications of Computer Vision 2019.
11 OSF 2018
This method uses stereo information.
 code 5.38 % 17.61 % 7.41 % 100.00 % 390 s 1 core @ 2.5 Ghz (Matlab + C/C++)
M. Menze, C. Heipke and A. Geiger: Object Scene Flow. ISPRS Journal of Photogrammetry and Remote Sensing (JPRS) 2018.
12 LiteFlowNet2-MD 7.36 % 7.79 % 7.43 % 100.00 % 0.058 s NVIDIA GTX 1080
13 LiteFlowNet2 code 7.62 % 7.64 % 7.62 % 100.00 % 0.0486 s GTX 1080 (slower than Pascal Titan X)
T. Hui, X. Tang and C. Loy: A Lightweight Optical Flow CNN - Revisiting Data Fidelity and Regularization. arXiv preprint arXiv:1903.07414 2019.
14 SENSE
This method uses stereo information.
 7.30 % 9.33 % 7.64 % 100.00 % 0.35 s GPU, GTX 1080Ti
15 IRR-PWC code 7.68 % 7.52 % 7.65 % 100.00 % 0.18 s NVIDIA GTX 1080 Ti
J. Hur and S. Roth: Iterative Residual Refinement for Joint Optical Flow and Occlusion Estimation. CVPR 2019.
16 CompactFlow 6.94 % 11.28 % 7.66 % 100.00 % 0.05 s Nvidia Tesla V100 (Python)
17 sceneflow1.0
This method uses stereo information.
 7.28 % 9.77 % 7.70 % 100.00 % 5 s GPU @ 2.5 Ghz (Python + C/C++)
18 PWC-Net+ code 7.69 % 7.88 % 7.72 % 100.00 % 0.03 s NVIDIA Pascal Titan X
D. Sun, X. Yang, M. Liu and J. Kautz: Models Matter, So Does Training: An Empirical Study of CNNs for Optical Flow Estimation. arXiv preprint arXiv:1809.05571 2018.
19 OSF
This method uses stereo information.
 code 5.62 % 18.92 % 7.83 % 100.00 % 50 min 1 core @ 2.5 Ghz (C/C++)
M. Menze and A. Geiger: Object Scene Flow for Autonomous Vehicles. Conference on Computer Vision and Pattern Recognition (CVPR) 2015.
20 oosf
This method uses stereo information.
 7.67 % 10.46 % 8.14 % 100.00 % 5 s GPU @ 2.5 Ghz (Python + C/C++)
21 CompactFlow_mix 7.37 % 13.45 % 8.38 % 100.00 % 0.05 s GPU @ 2.5 Ghz (Python)
22 SelFlow
This method makes use of multiple (>2) views.
 7.61 % 12.48 % 8.42 % 100.00 % 0.09 s GPU @ 2.5 Ghz (Python)
P. Liu, M. Lyu, I. King and J. Xu: SelFlow: Self-Supervised Learning of Optical Flow. CVPR 2019.
23 TEMP 8.82 % 7.73 % 8.64 % 100.00 %
24 CompactFlow_ROB 7.78 % 15.03 % 8.99 % 100.00 % 0.05 s 1 core @ 2.5 Ghz (Python)
25 SPN 9.00 % 9.87 % 9.15 % 100.00 % 0.08 s GPU @ 2.5 Ghz (C/C++)
26 FDFlowNet 9.31 % 9.71 % 9.38 % 100.00 % 0.02 s GPU @ 2.5 Ghz (Python)
27 LiteFlowNet code 9.66 % 7.99 % 9.38 % 100.00 % 0.0885 s GTX 1080 (slower than Pascal Titan X)
T. Hui, X. Tang and C. Loy: LiteFlowNet: A Lightweight Convolutional Neural Network for Optical Flow Estimation. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2018.
28 PWC-Net code 9.66 % 9.31 % 9.60 % 100.00 % 0.03 s NVIDIA Pascal Titan X
D. Sun, X. Yang, M. Liu and J. Kautz: PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume. CVPR 2018.
29 HTC
This method uses stereo information.
 9.15 % 12.76 % 9.75 % 100.00 % 0.03 s 1 core @ 2.5 Ghz (C/C++)
30 ContinualFlow_ROB
This method makes use of multiple (>2) views.
 8.54 % 17.48 % 10.03 % 100.00 % 0.15 s GPU - NVidia 1080Ti
M. Neoral, J. Šochman and J. Matas: Continual Occlusions and Optical Flow Estimation. 14th Asian Conference on Computer Vision (ACCV) 2018.
31 MirrorFlow code 8.93 % 17.07 % 10.29 % 100.00 % 11 min 4 core @ 2.2 Ghz (C/C++)
J. Hur and S. Roth: MirrorFlow: Exploiting Symmetries in Joint Optical Flow and Occlusion Estimation. ICCV 2017.
32 FlowNet2 code 10.75 % 8.75 % 10.41 % 100.00 % 0.1 s GPU @ 2.5 Ghz (C/C++)
E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovitskiy and T. Brox: FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017.
33 SS-SF
This method uses stereo information.
 8.17 % 25.20 % 11.00 % 100.00 % 3 min 1 core @ 2.5 Ghz (Matlab + C/C++)
34 SDF 8.61 % 23.01 % 11.01 % 100.00 % TBA 1 core @ 2.5 Ghz (C/C++)
M. Bai*, W. Luo*, K. Kundu and R. Urtasun: Exploiting Semantic Information and Deep Matching for Optical Flow. ECCV 2016.
35 FlowSubmission2015 9.99 % 16.67 % 11.10 % 100.00 % 0.06 s GPU @ 2.5 Ghz (Python)
36 UnFlow code 10.15 % 15.93 % 11.11 % 100.00 % 0.12 s GPU @ 1.5 Ghz (Python + C/C++)
S. Meister, J. Hur and S. Roth: UnFlow: Unsupervised Learning of Optical Flow with a Bidirectional Census Loss. AAAI 2018.
37 FSF+MS
This method uses stereo information.
This method makes use of the epipolar geometry.
This method makes use of multiple (>2) views.
 8.48 % 25.43 % 11.30 % 100.00 % 2.7 s 4 cores @ 3.5 Ghz (C/C++)
T. Taniai, S. Sinha and Y. Sato: Fast Multi-frame Stereo Scene Flow with Motion Segmentation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2017) 2017.
38 Flow2015 10.09 % 18.48 % 11.49 % 100.00 % 0.05 s 1 core @ 2.5 Ghz (Python)
39 CNNF+PMBP 10.08 % 18.56 % 11.49 % 100.00 % 45 min 1 cores @ 3.5 Ghz (C/C++)
F. Zhang and B. Wah: Fundamental Principles on Learning New Features for Effective Dense Matching. IEEE Transactions on Image Processing 2018.
40 stflow-ft 11.75 % 10.55 % 11.55 % 100.00 % 0.05 s 1 core @ 2.5 Ghz (C/C++)
41 PWC-Net_ROB code 11.22 % 13.69 % 11.63 % 100.00 % 0.03 s NVIDIA Pascal Titan X
D. Sun, X. Yang, M. Liu and J. Kautz: PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume. CVPR 2018.
42 SFF++
This method uses stereo information.
This method makes use of multiple (>2) views.
 10.63 % 17.48 % 11.77 % 100.00 % 78 s 4 cores @ 3.5 Ghz (C/C++)
43 SfM-PM
This method makes use of multiple (>2) views.
 9.66 % 22.73 % 11.83 % 100.00 % 69 s 3 cores @ 3.6 Ghz (C/C++)
D. Maurer, N. Marniok, B. Goldluecke and A. Bruhn: Structure-from-Motion-Aware PatchMatch for Adaptive Optical Flow Estimation. ECCV 2018.
44 DWARF
This method uses stereo information.
 11.88 % 12.15 % 11.93 % 100.00 % 0.09s - 1.43s TitanXP - JetsonTX2
45 MR-Flow
This method makes use of multiple (>2) views.
 code 10.13 % 22.51 % 12.19 % 100.00 % 8 min 1 core @ 2.5 Ghz (Python + C/C++)
J. Wulff, L. Sevilla-Lara and M. Black: Optical Flow in Mostly Rigid Scenes. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) 2017.
46 semflow 11.49 % 18.12 % 12.60 % 100.00 % 10 s 6 cores @ 2.5 Ghz (C/C++)
47 Mono-SF
This method uses stereo information.
 11.40 % 19.64 % 12.77 % 100.00 % 41 s 1 core @ 3.5 Ghz (Matlab + C/C++)
F. Brickwedde, S. Abraham and R. Mester: Mono-SF: Multi-View Geometry meets Single-View Depth for Monocular Scene Flow Estimation of Dynamic Traffic Scenes. accepted to International Conference on Computer Vision (ICCV) 2019.
48 SceneFFields
This method uses stereo information.
 10.58 % 24.41 % 12.88 % 100.00 % 65 s 4 cores @ 3.7 Ghz (C/C++)
R. Schuster, O. Wasenmüller, G. Kuschk, C. Bailer and D. Stricker: SceneFlowFields: Dense Interpolation of Sparse Scene Flow Correspondences. IEEE Winter Conference on Applications of Computer Vision (WACV) 2018.
49 FrequencyFlow+SegPM code 13.08 % 12.18 % 12.93 % 100.00 % 125 s 1 core @ >3.5 Ghz (C/C++)
50 CSF
This method uses stereo information.
 10.40 % 25.78 % 12.96 % 100.00 % 80 s 1 core @ 2.5 Ghz (C/C++)
Z. Lv, C. Beall, P. Alcantarilla, F. Li, Z. Kira and F. Dellaert: A Continuous Optimization Approach for Efficient and Accurate Scene Flow. European Conf. on Computer Vision (ECCV) 2016.
51 PWOC-3D
This method uses stereo information.
 12.40 % 15.78 % 12.96 % 100.00 % 0.13 s GTX 1080 Ti
R. Saxena, R. Schuster, O. Wasenmüller and D. Stricker: PWOC-3D: Deep Occlusion-Aware End-to-End Scene Flow Estimation. Intelligent Vehicles Symposium (IV) 2019.
52 PR-Sceneflow
This method uses stereo information.
 code 11.73 % 24.33 % 13.83 % 100.00 % 150 s 4 core @ 3.0 Ghz (Matlab + C/C++)
C. Vogel, K. Schindler and S. Roth: Piecewise Rigid Scene Flow. ICCV 2013.
53 SelFlow
This method makes use of multiple (>2) views.
 12.68 % 21.74 % 14.19 % 100.00 % 0.09 s GPU @ 2.5 Ghz (Python)
P. Liu, M. Lyu, I. King and J. Xu: SelFlow: Self-Supervised Learning of Optical Flow. CVPR 2019.
54 DDFlow 13.08 % 20.40 % 14.29 % 100.00 % 0.06 s GPU @ >3.5 Ghz (Python + C/C++)
P. Liu, I. King, M. Lyu and J. Xu: DDFlow: Learning Optical Flow with Unlabeled Data Distillation. AAAI 2019.
55 Devon 13.28 % 19.49 % 14.31 % 100.00 % 0.04 s 1 core @ 2.5 Ghz (C/C++)
56 DCFlow code 13.10 % 23.70 % 14.86 % 100.00 % 8.6 s GPU @ 3.0 Ghz (Matlab + C/C++)
J. Xu, R. Ranftl and V. Koltun: Accurate Optical Flow via Direct Cost Volume Processing. CVPR 2017.
57 ProFlow
This method makes use of multiple (>2) views.
 13.86 % 20.91 % 15.04 % 100.00 % 112 s GPU+CPU @ 3.6 Ghz (Python + C/C++)
D. Maurer and A. Bruhn: ProFlow: Learning to Predict Optical Flow. BMVC 2018.
58 FlowNetC-MD 15.98 % 10.96 % 15.14 % 100.00 % TBD s 1 core @ 2.5 Ghz (C/C++)
59 FlowFields++ code 14.82 % 17.77 % 15.31 % 100.00 % 29 s 1 core @ 3.5 Ghz (C/C++)
R. Schuster, C. Bailer, O. Wasenmüller and D. Stricker: FlowFields++: Accurate Optical Flow Correspondences Meet Robust Interpolation. International Conference on Image Processing (ICIP) 2018.
60 ProFlow_ROB
This method makes use of multiple (>2) views.
 14.15 % 21.82 % 15.42 % 100.00 % 112 s GPU+CPU @ 3.6 Ghz (Python + C/C++)
D. Maurer and A. Bruhn: ProFlow: Learning to Predict Optical Flow. BMVC 2018.
61 Deep-EIP 13.40 % 26.51 % 15.58 % 100.00 % 0.2 s GPU @ 2.5 Ghz (Python + C/C++)
62 HCF 14.95 % 19.37 % 15.68 % 100.00 % 0.02 s 1 core @ 2.5 Ghz (C/C++)
63 FF++_ROB 15.32 % 19.27 % 15.97 % 100.00 % 29 s 1 core @ 3.5 Ghz (C/C++)
R. Schuster, C. Bailer, O. Wasenmüller and D. Stricker: FlowFields++: Accurate Optical Flow Correspondences Meet Robust Interpolation. International Conference on Image Processing (ICIP) 2018.
64 SOF code 14.63 % 22.83 % 15.99 % 100.00 % 6 min 1 core @ 2.5 Ghz (Matlab)
L. Sevilla-Lara, D. Sun, V. Jampani and M. Black: Optical Flow with Semantic Segmentation and Localized Layers. CVPR 2016.
65 EPIflow 15.18 % 21.54 % 16.24 % 100.00 % 0.1 s GPU @ 2.0 Ghz (Python)
66 DIP-Flow-DF
This method makes use of multiple (>2) views.
 14.93 % 23.37 % 16.33 % 100.00 % 104s 2 cores @ 3.6 Ghz (C/C++)
D. Maurer, M. Stoll and A. Bruhn: Directional Priors for Multi-Frame Optical Flow. BMVC 2018.
67 JFS
This method makes use of the epipolar geometry.
 15.90 % 19.31 % 16.47 % 100.00 % 13 min 1 core @ 3.2 Ghz (C/C++)
J. Hur and S. Roth: Joint Optical Flow and Temporally Consistent Semantic Segmentation. ECCV Workshops 2016.
68 DF+OIR 15.11 % 23.45 % 16.50 % 100.00 % 3 min 1 core @ 3.5 Ghz (Matlab + C/C++)
D. Maurer, M. Stoll and A. Bruhn: Order-Adaptive and Illumination Aware Variational Optical Flow Refinement. BMVC 2017.
69 SPS+FF++
This method uses stereo information.
 code 15.91 % 20.27 % 16.64 % 100.00 % 36 s 1 core @ 3.5 Ghz (C/C++)
R. Schuster, O. Wasenmüller and D. Stricker: Dense Scene Flow from Stereo Disparity and Optical Flow. ACM Computer Science in Cars Symposium (CSCS) 2018.
70 DIP-Flow-CPM
This method makes use of multiple (>2) views.
 15.57 % 23.84 % 16.95 % 100.00 % 52 s 2 core @ 3.6 Ghz (C/C++)
D. Maurer, M. Stoll and A. Bruhn: Directional Priors for Multi-Frame Optical Flow. BMVC 2018.
71 ImpPB+SPCI code 17.25 % 20.44 % 17.78 % 100.00 % 60 s GPU @ 2.5 Ghz (Python)
T. Schuster, L. Wolf and D. Gadot: Optical Flow Requires Multiple Strategies (but only one network). CVPR 2017.
72 MPIF 17.72 % 18.97 % 17.93 % 100.00 % 5 s 1 core @ 3.5 Ghz (C/C++)
73 UnOS(Full)
This method uses stereo information.
 16.93 % 23.34 % 18.00 % 100.00 % 0.08 s 1 core @ 2.5 Ghz (C/C++)
74 PCOF-LDOF
This method uses stereo information.
 14.34 % 38.32 % 18.33 % 100.00 % 50 s 1 core @ 3.0 Ghz (C/C++)
M. Derome, A. Plyer, M. Sanfourche and G. Le Besnerais: A Prediction-Correction Approach for Real-Time Optical Flow Computation Using Stereo. German Conference on Pattern Recognition 2016.
75 UFlow 16.86 % 26.46 % 18.46 % 100.00 % 0.02 s GPU @ 3.0 Ghz (Python)
76 FlowFieldCNN 18.33 % 20.42 % 18.68 % 100.00 % 23 s GPU/CPU 4 core @ 3.5 Ghz (C/C++)
C. Bailer, K. Varanasi and D. Stricker: CNN-based Patch Matching for Optical Flow with Thresholded Hinge Embedding Loss. CVPR 2017.
77 RicFlow 18.73 % 19.09 % 18.79 % 100.00 % 5 s 1 core @ 3.5 Ghz (C/C++)
Y. Hu, Y. Li and R. Song: Robust Interpolation of Correspondences for Large Displacement Optical Flow. CVPR 2017.
78 EAI-Flow 18.05 % 26.23 % 19.41 % 100.00 % 4.5 s 2 cores @ 3.0 Ghz (Python + C/C++)
79 PGM-G 18.90 % 23.43 % 19.66 % 100.00 % 5.05 s 1 core @ 3.1 Ghz (C/C++)
Y. Li: Pyramidal Gradient Matching for Optical Flow Estimation. CoRR 2017.
80 FlowFields+ 19.51 % 21.26 % 19.80 % 100.00 % 28s 1 core @ 3.5 Ghz (C/C++)
C. Bailer, B. Taetz and D. Stricker: Flow Fields: Dense Correspondence Fields for Highly Accurate Large Displacement Optical Flow Estimation. .
81 PMPM code 19.36 % 22.19 % 19.83 % 100.00 % 1.1 s 1 core @ 3.5 Ghz (C/C++)
82 GeoFlow 19.67 % 20.85 % 19.87 % 100.00 % 0.04 s 1 core @ 2.5 Ghz (C/C++)
83 joint flow 18.34 % 29.14 % 20.14 % 100.00 % 0.08 s GPU @ 2.5 Ghz (C/C++)
84 EPC++ (stereo)
This method uses stereo information.
 19.24 % 26.93 % 20.52 % 100.00 % 0.05 s GPU @ 2.5 Ghz (Python)
C. Luo, Z. Yang, P. Wang, Y. Wang, W. Xu, R. Nevatia and A. Yuille: Every pixel counts++: Joint learning of geometry and motion with 3D holistic understanding. arXiv preprint arXiv:1810.06125 2018.
85 PatchBatch code 19.98 % 26.50 % 21.07 % 100.00 % 50 s GPU @ 2.5 Ghz (Python)
D. Gadot and L. Wolf: PatchBatch: a Batch Augmented Loss for Optical Flow. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2016.
86 DDF code 20.36 % 25.19 % 21.17 % 100.00 % ~1 min GPU @ 2.5 Ghz (C/C++)
F. G\"uney and A. Geiger: Deep Discrete Flow. Asian Conference on Computer Vision (ACCV) 2016.
87 SODA-Flow 20.01 % 29.14 % 21.53 % 100.00 % 96 s 2 cores @ 3.5 Ghz (C/C++)
D. Maurer, M. Stoll, S. Volz, P. Gairing and A. Bruhn: A Comparison of Isotropic and Anisotropic Second Order Regularisers for Optical Flow. SSVM 2017.
88 boostflow 20.74 % 25.57 % 21.54 % 100.00 % 0.08 s GPU @ 2.5 Ghz (C/C++)
89 EPC++ 20.61 % 26.32 % 21.56 % 100.00 % 0.05 s GPU @ 2.5 Ghz (Python)
90 DiscreteFlow code 21.53 % 21.76 % 21.57 % 100.00 % 3 min 1 core @ 2.5 Ghz (Matlab + C/C++)
M. Menze, C. Heipke and A. Geiger: Discrete Optimization for Optical Flow. German Conference on Pattern Recognition (GCPR) 2015.
91 SGM+SF
This method uses stereo information.
 20.91 % 25.50 % 21.67 % 100.00 % 45 min 16 core @ 3.2 Ghz (C/C++)
H. Hirschmüller: Stereo Processing by Semiglobal Matching and Mutual Information. PAMI 2008.
M. Hornacek, A. Fitzgibbon and C. Rother: SphereFlow: 6 DoF Scene Flow from RGB-D Pairs. CVPR 2014.
92 OAR-Flow 20.62 % 27.67 % 21.79 % 100.00 % 100 s 2 cores @ 3.5 Ghz (C/C++)
D. Maurer, M. Stoll and A. Bruhn: Order-Adaptive Regularisation for Variational Optical Flow: Global, Local and in Between. SSVM 2017.
93 OD-Flow 20.89 % 28.03 % 22.08 % 100.00 % 0.3 s GPU @ 3.5 Ghz (Python)
94 CPM-Flow code 22.32 % 22.81 % 22.40 % 100.00 % 4.2 s 1 core @ 3.5 Ghz (C/C++)
Y. Hu, R. Song and Y. Li: Efficient Coarse-to-Fine PatchMatch for Large Displacement Optical Flow. CVPR 2016.
95 PCOF + ACTF
This method uses stereo information.
 14.89 % 60.15 % 22.43 % 100.00 % 0.08 s GPU @ 2.0 Ghz (C/C++)
M. Derome, A. Plyer, M. Sanfourche and G. Le Besnerais: A Prediction-Correction Approach for Real-Time Optical Flow Computation Using Stereo. German Conference on Pattern Recognition 2016.
96 SegFlow(d0=3) 22.21 % 23.72 % 22.46 % 100.00 % 6.6 s 1 core @ >3.5 Ghz (C/C++)
J. Chen, Z. Cai, J. Lai and X. Xie: Efficient Segmentation-based PatchMatch for Large displacement Optical Flow Estimation. IEEE TCSVT 2018.
97 IntrpNt-df code 22.15 % 26.03 % 22.80 % 100.00 % 3 min GPU @ 2.5 Ghz (Python)
S. Zweig and L. Wolf: InterpoNet, a Brain Inspired Neural Network for Optical Flow Dense Interpolation. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017.
98 SGM&FlowFie+
This method uses stereo information.
 22.83 % 22.75 % 22.82 % 81.24 % 29 s 1 core @ 3.5 Ghz (C/C++)
R. Schuster, C. Bailer, O. Wasenmüller and D. Stricker: Combining Stereo Disparity and Optical Flow for Basic Scene Flow. Commercial Vehicle Technology Symposium (CVTS) 2018.
99 Back2FutureFlow
This method makes use of multiple (>2) views.
 code 22.67 % 24.27 % 22.94 % 100.00 % 0.12 s GPU @ 2.5 Ghz (LUA/Torch)
J. Janai, F. Güney, A. Ranjan, M. Black and A. Geiger: Unsupervised Learning of Multi-Frame Optical Flow with Occlusions. Proc. of the European Conf. on Computer Vision (ECCV) 2018.
100 MotionSLIC
This method makes use of the epipolar geometry.
 code 14.86 % 64.44 % 23.11 % 100.00 % 30 s 4 cores @ 2.5 Ghz (C/C++)
K. Yamaguchi, D. McAllester and R. Urtasun: Robust Monocular Epipolar Flow Estimation. CVPR 2013.
101 IntrpNt-cpm code 22.51 % 26.54 % 23.18 % 100.00 % 5.6 s GPU @ 2.5 Ghz (Python)
S. Zweig and L. Wolf: InterpoNet, a Brain Inspired Neural Network for Optical Flow Dense Interpolation. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017.
102 FullFlow 23.09 % 24.79 % 23.37 % 100.00 % 4 min 4 cores @ >3.5 Ghz (Matlab and C++)
Q. Chen and V. Koltun: Full Flow: Optical Flow Estimation By Global Optimization over Regular Grids. CVPR 2016.
103 HiLM code 23.73 % 21.79 % 23.41 % 100.00 % 8 sec P6000 (C/C++)
M. Fathy, Q. Tran, M. Zia, P. Vernaza and M. Chandraker: Hierarchical Metric Learning and Matching for 2D and 3D Geometric Correspondences. European Conference on Computer Vision (ECCV) 2018.
104 PatchFlow 23.52 % 23.14 % 23.46 % 100.00 % 0.08 s 1 core @ 2.5 Ghz (C/C++)
105 boostflow_fnet 22.69 % 28.10 % 23.59 % 100.00 % 0.08 s GPU @ 2.5 Ghz (C/C++)
106 UnOS(flownet-only) 23.49 % 25.05 % 23.75 % 100.00 % 0.08 s 1 core @ 2.5 Ghz (C/C++)
107 IntrpNt-dm code 23.46 % 26.27 % 23.93 % 100.00 % 15 s GPU @ 2.5 Ghz (Python)
S. Zweig and L. Wolf: InterpoNet, a Brain Inspired Neural Network for Optical Flow Dense Interpolation. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017.
108 SPM-BP 24.06 % 24.97 % 24.21 % 100.00 % 10 s 2 cores @ 2.5 Ghz (C/C++)
Y. Li, D. Min, M. Brown, M. Do and J. Lu: SPM-BP: Sped-up PatchMatch Belief Propagation for Continuous MRFs. Proceedings of the IEEE International Conference on Computer Vision 2015.
109 vsv 24.66 % 25.24 % 24.75 % 100.00 % 124 s 1 core @ >3.5 Ghz (Matlab + C/C++)
110 CC-Flow 25.10 % 26.11 % 25.27 % 100.00 % 0.06 s GPU @ 2.5 Ghz (Python)
111 PPM code 25.87 % 23.67 % 25.50 % 100.00 % 17.3 s 1 core @ 2.5 Ghz (C/Chttps://github.c++)
F. Kuang: PatchMatch algorithms for motion estimation and stereo reconstruction. 2017.
112 3DFlow 25.56 % 29.33 % 26.19 % 100.00 % 448s Matlab with embedded C++ code
J. Chen, Z. Cai, J. Lai and X. Xie: A Filtering Based Framework for Optical Flow Estimation. IEEE TCSVT 2018.
113 EpicFlow code 25.81 % 28.69 % 26.29 % 100.00 % 15 s 1 core @ >3.5 Ghz (C/C++)
J. Revaud, P. Weinzaepfel, Z. Harchaoui and C. Schmid: EpicFlow: Edge-Preserving Interpolation of Correspondences for Optical Flow. CVPR 2015 - IEEE Conference on Computer Vision \& Pattern Recognition 2015.
114 SegFlow(d0=11) 28.97 % 22.64 % 27.91 % 100.00 % 4.5 s 1 core @ 3.5 Ghz (C/C++)
J. Chen, Z. Cai, J. Lai and X. Xie: Efficient Segmentation-based PatchMatch for Large displacement Optical Flow Estimation. IEEE TCSVT 2018.
115 OD-Flow 27.48 % 33.26 % 28.44 % 100.00 % 0.3s s GPU @ 3.5 Ghz (Python)
116 DeepFlow code 27.96 % 31.06 % 28.48 % 100.00 % 17 s 1 core @ >3.5 Ghz (Python + C/C++)
P. Weinzaepfel, J. Revaud, Z. Harchaoui and C. Schmid: DeepFlow: Large displacement optical flow with deep matching. IEEE Intenational Conference on Computer Vision (ICCV) 2013.
117 OD-Flow 27.81 % 32.24 % 28.54 % 100.00 % 0.3 s GPU @ 3.5 Ghz (Python)
118 DeGraF-Flow 28.78 % 29.69 % 28.94 % 100.00 % 3.2 s 4 cores @ 2.5 Ghz (C/C++)
119 CPNFlow 31.05 % 27.16 % 30.40 % 100.00 % 0.1 s GPU @ 1.5 Ghz (Python)
Y. Yang and S. Soatto: Conditional prior networks for optical flow. Proceedings of the European Conference on Computer Vision (ECCV) 2018.
120 IIOF-NLDP 30.23 % 32.44 % 30.60 % 100.00 % 350 s 4 cores @ 3.5 Ghz (Matlab + C/C++)
D. Trinh, W. Blondel and C. Daul: A General Form of Illumination- Invariant Descriptors in Variational Optical Flow Estimation. IEEE Int. Conf. on Image Processing (ICIP) 2017.
121 DMF_ROB code 30.74 % 30.07 % 30.63 % 100.00 % 150 s 1 core @ 2.5 Ghz (C/C++)
P. Weinzaepfel, J. Revaud, Z. Harchaoui and C. Schmid: DeepFlow: Large displacement optical flow with deep matching. ICCV - IEEE International Conference on Computer Vision 2013.
122 SPyNet code 33.36 % 43.62 % 35.07 % 100.00 % 0.16 s 1 core @ 2.5 Ghz (C/C++)
A. Ranjan and M. Black: Optical Flow Estimation using a Spatial Pyramid Network. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2017.
123 SGM+C+NL
This method uses stereo information.
 code 34.24 % 42.46 % 35.61 % 93.83 % 4.5 min 1 core @ 2.5 Ghz (C/C++)
H. Hirschmüller: Stereo Processing by Semiglobal Matching and Mutual Information. PAMI 2008.
D. Sun, S. Roth and M. Black: A Quantitative Analysis of Current Practices in Optical Flow Estimation and the Principles Behind Them. IJCV 2013.
124 DWBSF
This method uses stereo information.
 40.74 % 31.16 % 39.14 % 100.00 % 7 min 4 cores @ 3.5 Ghz (C/C++)
C. Richardt, H. Kim, L. Valgaerts and C. Theobalt: Dense Wide-Baseline Scene Flow From Two Handheld Video Cameras. 3DV 2016.
125 SGM+LDOF
This method uses stereo information.
 code 40.81 % 31.92 % 39.33 % 95.89 % 86 s 1 core @ 2.5 Ghz (C/C++)
H. Hirschmüller: Stereo Processing by Semiglobal Matching and Mutual Information. PAMI 2008.
T. Brox and J. Malik: Large Displacement Optical Flow: Descriptor Matching in Variational Motion Estimation. PAMI 2011.
126 HS code 39.90 % 51.39 % 41.81 % 100.00 % 2.6 min 1 core @ 3.0 Ghz (Matlab)
D. Sun, S. Roth and M. Black: A Quantitative Analysis of Current Practices in Optical Flow Estimation and The Principles Behind Them. 2014.
127 UnOS(Ego) 34.22 % 92.15 % 43.86 % 100.00 % 0.08 s 1 core @ 2.5 Ghz (C/C++)
128 GCSF
This method uses stereo information.
 code 47.38 % 41.50 % 46.40 % 100.00 % 2.4 s 1 core @ 2.5 Ghz (C/C++)
J. Cech, J. Sanchez-Riera and R. Horaud: Scene Flow Estimation by growing Correspondence Seeds. CVPR 2011.
129 DB-TV-L1 code 47.52 % 48.27 % 47.64 % 100.00 % 16 s 1 core @ 2.5 Ghz (Matlab)
C. Zach, T. Pock and H. Bischof: A Duality Based Approach for Realtime TV- L1 Optical Flow. DAGM 2007.
130 VSF
This method uses stereo information.
 code 50.06 % 45.40 % 49.28 % 100.00 % 125 min 1 core @ 2.5 Ghz (C/C++)
F. Huguet and F. Devernay: A Variational Method for Scene Flow Estimation from Stereo Sequences. ICCV 2007.
131 HAOF code 49.89 % 50.74 % 50.04 % 100.00 % 16.2 s 1 core @ 2.5 Ghz (C/C++)
T. Brox, A. Bruhn, N. Papenberg and J. Weickert: High accuracy optical flow estimation based on a theory for warping. ECCV 2004.
132 TVL1_ROB code 51.15 % 51.12 % 51.14 % 100.00 % 3 s 4 cores @ 2.5 Ghz (C/C++)
J. Sánchez Pérez, E. Meinhardt-Llopis and G. Facciolo: TV-L1 Optical Flow Estimation. Image Processing On Line 2013.
133 PolyExpand 52.00 % 58.56 % 53.09 % 100.00 % 1 s 1 core @ 2.5 Ghz (C/C++)
G. Farneback: Two-Frame Motion Estimation Based on Polynomial Expansion. SCIA 2003.
134 SPFlow 55.28 % 67.18 % 57.26 % 100.00 % 10 s 1 core @ 2.5 Ghz (Matlab)
135 PMSFlow 62.09 % 61.80 % 62.04 % 100.00 % 0.08 s 1 core @ 2.5 Ghz (C/C++)
136 H+S_ROB code 68.22 % 76.49 % 69.60 % 100.00 % 8 s 4 cores @ 2.5 Ghz (C/C++)
E. Meinhardt-Llopis, J. Sánchez Pérez and D. Kondermann: Horn-Schunck Optical Flow with a Multi-Scale Strategy. Image Processing On Line 2013.
137 Pyramid-LK code 71.84 % 76.82 % 72.67 % 100.00 % 1.5 min 1 core @ 2.5 Ghz (Matlab)
J. Bouguet: Pyramidal implementation of the Lucas Kanade feature tracker. Intel 2000.
138 MEDIAN 87.37 % 92.80 % 88.27 % 99.86 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
139 AVERAGE 88.47 % 92.08 % 89.07 % 99.86 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
140 AVG_FLOW_ROB 89.48 % 92.60 % 90.00 % 100.00 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
141 UFT 96.10 % 97.07 % 96.26 % 1.55 % 0.8 s 1 core @ 2.5 Ghz (C/C++)
Table as LaTeX | Only published Methods




Related Datasets

  • HCI/Bosch Robust Vision Challenge: Optical flow and stereo vision challenge on high resolution imagery recorded at a high frame rate under diverse weather conditions (e.g., sunny, cloudy, rainy). The Robert Bosch AG provides a prize for the best performing method.
  • Image Sequence Analysis Test Site (EISATS): Synthetic image sequences with ground truth information provided by UoA and Daimler AG. Some of the images come with 3D range sensor information.
  • Middlebury Stereo Evaluation: The classic stereo evaluation benchmark, featuring four test images in version 2 of the benchmark, with very accurate ground truth from a structured light system. 38 image pairs are provided in total.
  • Daimler Stereo Dataset: Stereo bad weather highway scenes with partial ground truth for freespace
  • Make3D Range Image Data: Images with small-resolution ground truth used to learn and evaluate depth from single monocular images.
  • Lubor Ladicky's Stereo Dataset: Stereo Images with manually labeled ground truth based on polygonal areas.
  • Middlebury Optical Flow Evaluation: The classic optical flow evaluation benchmark, featuring eight test images, with very accurate ground truth from a shape from UV light pattern system. 24 image pairs are provided in total.

Citation

When using this dataset in your research, we will be happy if you cite us:
@ARTICLE{Menze2018JPRS,
  author = {Moritz Menze and Christian Heipke and Andreas Geiger},
  title = {Object Scene Flow},
  journal = {ISPRS Journal of Photogrammetry and Remote Sensing (JPRS)},
  year = {2018}
}
@INPROCEEDINGS{Menze2015ISA,
  author = {Moritz Menze and Christian Heipke and Andreas Geiger},
  title = {Joint 3D Estimation of Vehicles and Scene Flow},
  booktitle = {ISPRS Workshop on Image Sequence Analysis (ISA)},
  year = {2015}
}



eXTReMe Tracker