Robust Optical Flow

GCPR 2013 - Special Session on Robust Optical Flow

Aims and Scope

The goal of this special session at the German Conference on Pattern Recognition (GCPR, formerly: DAGM) 2013 is to foster research on robust optical flow with a focus on challenging real-world scenes in the KITTI optical flow benchmark. While many optical flow methods have been published over the last decades, little attention has been paid to realistic scenarios as they arise for example in autonomous driving or blind people assistance applications. Here, typical difficulties include illumination changes and large displacements, but also low-textured areas, reflections and specularities. Papers submitted to this special session are not required to address all problems, but to provide insights or solutions for at least one of these problems. For comparison, all submissions are required to provide results on the same set of images.

Download: (Call for papers (pdf)

                                    Illumination Changes                                                              Large Displacements


Topics of interest include, but are not limited to:

  • Robust models
  • Inclusion of high-level knowledge
  • Detection of model violations
  • Automatic model selection
  • Confidence estimation
  • Features
  • Performance analysis
  • Literature survey papers
  • Evaluation of existing methods

Dataset / Downloads

Papers submitted to this special session should address at least one of the following two problems. Each submission is expected to show results on the four images provided for each problem (Note: In order to make ground truth available, these images are taken from the KITTI optical flow training set and used as test images for this special session):

Additionally, all authors are encouraged to evaluate their method on the full KITTI optical flow dataset, submit their results to the evaluation server and discuss their findings in the paper. Note that we also provide a MATLAB / C++ development kit for reading and writing optical flow maps (flow_read.m, flow_write.m), including a readme documenting the data format and tools for generating optical flow error images and histograms for comparison (flow_error_image.m, flow_error_histogram.m).

Baseline Results

Using state-of-the-art optical flow algorithms, we have created some baseline results on the eight test images to give better insights into the problems of current methods. The numbers are bad pixels (BP in percent, error threshold = 3 px) and average endpoint error (AEE in pixels), computed over all pixels for which ground truth is available (including occluded surfaces, i.e., 'ooc'). Error images and histograms are available by clicking at the respective link. Implementations for computing these measures and visualizations are available in the development kit. All baseline results can also be downloaded in a single zip file.

Illumination Changes:

Method#44#11#15#74Avg. BP
OFH 201123.20 % (5.11 Px)37.25 % (12.47 Px)32.19 % (9.05 Px)62.90 % (24.00 Px)38.89 %
SRB 201026.58 % (4.67 Px)40.60 % (13.76 Px)32.84 % (9.72 Px)62.93 % (24.27 Px)40.74 %
SRBF 201031.82 % (5.62 Px)40.33 % (13.96 Px)35.12 % (12.40 Px)64.88 % (24.64 Px)43.04 %
BW 2005 32.42 % (5.19 Px)33.93 % (8.50 Px)47.69 % (12.17 Px)71.43 % (25.14 Px)46.73 %
HS 198142.92 % (6.77 Px)38.83 % (10.72 Px)58.08 % (12.89 Px)82.14 % (28.75 Px)55.94 %
WPB 201049.08 % (9.20 Px)49.99 % (28.36 Px)67.27 % (28.36 Px)88.66 % (30.68 Px)63.75 %

Large Displacements:

Method#147#117#144#181Avg. BP
OFH 201115.03 % (4.69 Px)16.26 % (4.33 Px)42.03 % (15.00 Px)63.58 % (50.51 Px)34.30 %
SRB 201014.58 % (4.84 Px)24.70 % (9.74 Px)50.68 % (19.02 Px)67.10 % (47.70 Px)39.27 %
SRBF 201014.78 % (5.16 Px)24.40 % (9.92 Px)50.66 % (19.34 Px)68.40 % (48.81 Px)39.56 %
BW 2005 16.99 % (5.16 Px)28.79 % (7.86 Px)46.98 % (16.84 Px)69.03 % (45.26 Px)40.45 %
HS 198124.84 % (6.60 Px)43.25 % (15.31 Px)51.88 % (14.80 Px)74.09 % (49.27 Px)48.52 %
WPB 201032.71 % (8.09 Px)46.80 % (13.67 Px)52.24 % (17.94 Px)75.99 % (50.18 Px)51.94 %


Author Guidelines

Papers submitted to this "Special Session on Robust Optical Flow" are regular GCPR (formerly: DAGM) papers, enter the GCPR 2013 review process and will appear in the GCPR 2013 proceedings (Springer LNCS - Lecture Notes in Computer Science) once they are accepted. Please follow the submission guidelines for GCPR 2013 to submit your paper and select "Special Session on Robust Optical Flow" during the submission process. Please use the functions provided in the development kit for visualizing your results. Additionally, also send your results on all eight test images to the organizing committee, such that they can be visualized and compared at the special session's panel discussion.

Important Dates

See GCPR 2013 conference website.

Organizing Committee

If you have any questions, do not hesitate to direct your questions to the session organizers: