StereoWalker Dataset

Public release of the datasets used in StereoWalker. It contains two parts:

Part	Directory	Contents	Size
StereoUrban (training)	StereoUrban/	544 YouTube video IDs + camera-pose annotations	~310 MB
UVA robot test set	uva_test/	10 stereo sequences, 3,305 samples with images, poses, and scene-category labels	~530 MB

Part 1: StereoUrban — YouTube stereo walking videos

We do not redistribute the videos themselves. We release the YouTube video IDs (StereoUrban/video_ids.txt) together with the per-video camera-pose annotations we generated (StereoUrban/annotations/<video_id>.npy). Each video can be retrieved at https://www.youtube.com/watch?v=<video_id>.

Left/right eye views with the upcoming 10 s path in the current camera frame (5× real time) — the short-horizon relative supervision the annotations are designed for.

See StereoUrban/README.md for the annotation format and processing details.

Part 2: UVA robot test set

A real-robot stereo navigation dataset collected on a wheeled robot. 10 continuous segments, 3,305 stereo samples at 1 Hz, each with a 6-DoF pose and (for 9 of the 10 segments) per-sample scene-category labels (crowd, person close by, turn, action–target mismatch, crossing, other).

Stereo camera views with the robot trajectory drawn in sync (test segment 3, 16× real time).

See uva_test/README.md for the directory layout and JSON schema.

License

The annotations and the UVA test set are released under CC BY 4.0 — see LICENSE. The YouTube videos referenced in StereoUrban/video_ids.txt remain the property of their respective uploaders and are subject to YouTube's Terms of Service; this release contains only video identifiers and annotations derived by us.