L. Mehl, J. Schmalfuss, A. Jahedi, Y. Nalivayko, A. Bruhn — University of Stuttgart

Not logged in | Login | Create Account

Frequently Asked Questions

How does the registration work?

In order to submit to the benchmark, i.e. evaluate results on the Spring test split, you need to register. Please note that in order to download the dataset, no registration is needed. To register, first create an account and confirm your mail address by clicking on the link we will send to you. Afterwards, your account request will be verified by our team. Finally, after successful verification, you can submit results for evaluation.

How does the submission work?

1. Let your new and interesting approach predict results for the spring test split.
• You can find code to read/write .flo5 and .dsp5 files here.
• Use the same folder structure, file naming and file formats as in the dataset.
• For stereo use <rootdir>/####/disp1_{left|right}/disp1_{left|right}_####.dsp5
• For optical flow use <rootdir>/####/flow_{FW|BW}_{left|right}/flow_{FW|BW}_{left|right}_####.flo5
• For scene flow use the above plus <rootdir>/####/disp2_{FW|BW}_{left|right}/disp2_{FW|BW}_{left|right}_####.dsp5
2. Use the subsampling executable for submission file(s) generation.
• Download the subsampling executable(s).
• Usage: ./{flow|disp1|disp2}_subsampling <rootdir>. A hdf5 submission file is generated in the current directory.
• Use the disp1 or flow executables for the stereo or optical flow task; use all three for the scene flow task.
3. Upload to our benchmark webpage for evaluation.
• Upload your hdf5 file (for scene flow: your three hdf5 files) under submit.
• Please allow up to 1-2h until your result is evaluated.
• All results are by default set to private. You can later choose results to be private, public anonymous or public.

Which data formats do you provide?

Images and maps are given in png format. Disparity and optical flow files are given in HDF5 file format and named .dsp5 for disparity and .flo5 for optical flow. You can find reference code to read/write .flo5 and .dsp5 files here.

Do you also provide point clouds for the Spring dataset?

You can find code to transform Spring data to point 3D point clouds here.

How can I compute metric depth from disparity?

In general, depth Z is computed from disparity d through Z = fx * B / d, where fx is the focal length in pixels (given in intrinsics.txt) and B is the stereo camera baseline distance; for Spring this is always 0.065m. Please note that the Spring dataset encodes infinitely distant sky pixels as zero disparity, leading to infinite values when using the above formula.

Which vector length is encoded in the ground truth data?

While our ground truth files are given in 4K (double the spatial resolution per dimension), the ground truth vectors (disparities, optical flow) relate to the images in full HD resolution. So when using these, there should be no need to divide them by 2. See also the example data loader for Spring here.

Which (stereo) camera settings does Spring use?

Spring uses an orthoparallel stereo camera setting, i.e. two cameras parallely pointing into the same direction. The baseline distance between the cameras is 0.065m. Intrinsic camera parameters (available for train and test) are given per sequence in intrinsics.txt, extrinsic camera data / camera poses (available for train) are given in extrinsics.txt. Please note that in some scenes a camera zoom / change of the focal length is used, leading to different intrinsics per frame. We additionally provide metric camera focal distances in focaldistance.txt.

Are there nan values in the ground truth files?

Yes, very occasionally, there are nan values in the ground truth files, which arise from a bug in the Blender/cycles shading system. During evaluation, these pixels are ignored.

I have further questions!

Please use the contact link at the bottom of the page.