DSEC-Flow: Optical Flow Benchmark

The task is to estimate dense optical flow from event cameras. Global shutter cameras can be used as well but their use must be specified in the submission.

Details

• The predictions will be evaluated on the undistorted and rectified view of the left event camera on a subset of image timestamps on the test set.
• Only forward optical flow will be evaluated.
• If the global shutter cameras are used, their usage must be specified within the submitted report or paper.
• The evaluation will be performed on all pixels on which ground truth optical flow is available.

Metrics

The main metric used for the ranking is the end-point-error (EPE) of the optical flow on the test set of DSEC-Flow.

We also provide the following error metrics:

• NPE: 1-pixel-error, the percentage of ground truth pixels with optical flow magnitude error > N. N is either 1, 2 or 3.
• EPE: Endpoint error. The average of the L2-Norm of the optical flow error.
• AE: Angular error.

A detailed explanation of EPE and AE can be found in “A Database and Evaluation Methodology for Optical Flow“.

Submission Format

The submission will be evaluated on the predicted forward optical flow in the left event camera on a subset of image timestamps on the test set. Visit the submission format page for more details.

Report (mandatory)

The submission must eventually be accompanied by a

• a workshop paper (max 8 pages, 50 MB, pdf) if the project is published at a workshop that is part of a conference (e.g. ICRA, CVPR workshops).
• a paper (max 100 MB, pdf) if the project is published at a conference or journal.

In either case, specify the venue (competition, workshop, conference, or journal and year) or write “under review” in the “Additional information” field of the submission.

Note that you can still add a report/paper after your submission.

We will regularly delete submissions older than 6 months without a report or paper or the specifics to identify the venue. If your work has been under review for more than 6 months you must resubmit your work.

Note on Double-Blind Review

If your submitted work is under review in a double-blind process. Consider adding “anonymous” to the author and affiliation field for the submission (as well as “under review” in the “additional information” field). When the work is accepted, you can delete the submission and resubmit with the full information mentioned above.

Submission Policy

• We require that the submission uses the same parameter set for the whole test set. I.e. it is not allowed to use different hyperparameters for different sequences.
• It is not allowed to use test data in any way to train your model for your submission. E.g. fine-tuning your model with a photometric loss on the test set is prohibited.
• It is allowed to submit multiple results for a final ablation study for paper submissions.
• Use your institutional email address (e.g., .edu).

Naming Convention

You can add the following abbreviation in curly brackets after your method name to highlight self-imposed constraints:

• selfsup: The method has not been trained on any ground truth optical flow data, including 3rd-party datasets. Instead, the method has been trained via self-supervision.
• optim: The method is optimization-based and has never used ground truth data.
• sim2real: The method has been trained entirely with simulated data.

Example: E-RAFT [sim2real], would be E-RAFT trained only on simulated data.

If you have further suggestions, please contact Mathias.

Questions

Contact mgehrig@ifi.uzh.ch for inquiries regarding the benchmark. Issues and questions regarding the dataset should be discussed on GitHub instead.

Citation

Cite the following work when using this benchmark or DSEC-Flow in general:

@InProceedings{Gehrig3dv2021,
  author = {Mathias Gehrig and Mario Millh\"ausler and Daniel Gehrig and Davide Scaramuzza},
  title = {E-RAFT: Dense Optical Flow from Event Cameras},
  booktitle = {International Conference on 3D Vision (3DV)},
  year = {2021}
}