Data supporting the publication: On the Real-World Generalisability of Optical Flow Models
What's in it
This dataset supports research into the real-world generalisability of optical flow models, the computer vision task of estimating per-pixel motion correspondences between consecutive video frames. It addresses the scarcity of accurately annotated real-world optical flow data by combining three standardised sub-datasets, together comprising 8,204 frame pairs. FlowFactor (1,000 frame pairs) consists of hand-annotated correspondences collected specifically for this benchmark using a custom annotation tool. SlowFlow (3,447 frame pairs) and TAP-Flow (3,757 frame pairs) are repurposed from existing datasets, Slow Flow (Janai et al., CVPR 2017) and TAP-Vid (Doersch et al. NeurIPS 2022) respectively, restructured into a standardised, KITTI-2015-style format for consistent benchmarking. Together, these sub-datasets enable more accurate assessment of optical flow model performance under real-world conditions than existing synthetic or narrow-domain benchmarks allow.
Files and services
- DataCite REST API (JSON)api.datacite.orgOpen