Optical Flow Evaluation V1 (BlinkFlow)

We provide a large training dataset (33K frames) and an evaluation benchmark (4K frames) that contains sufficient, diversiform, and challenging event data with optical flow ground truth.

Download links:

• Download training data (244 GB)
• Download test data (25 GB)
• Download samples map (70 MB, used for submission)

We provide some scripts to help you submit your method. The order should be:

1. Download the test data and the sample maps.
2. Run the inference code to generate the flow result for your method.
3. Use the provided sample_data.py to pack the result.
4. Use the provided check_submission.py to check your submission.
5. Submit your result to the website.

The evaluation table shows the most commonly used metrics. To access more metrics for a given method, clink on the method name to go to the detail page. The metrics used in the table are:

• EPE: end-point error, i.e., L2-norm of the optical flow error.
• Outlier: EPE > 3 or error rate > 5% (error rate equals to epe divided by ground truth magnitude).
• AE: angular error.
• N-pe: the percentage that the EPE is larger than N, where N is 1, 2, 3, 5.

	Methods	EPE	Out	AE	1pe	2pe	3pe