kestrel dataset, public part

Video clips and precomputed RGB features. Every clip is named by a random UUID; meta.json at the bucket root lists them all.

files/<uuid>/video.mp4   source video (remuxed without re-encoding; container metadata removed)
files/<uuid>/rgb.npz     RGB features (numpy .npz, T = frames), absent for some clips
meta.json                clip index

rgb.npz arrays

Left/right are the person's anatomical left/right (COCO wrists 9/10).

Videos

Each video.mp4 is a stream copy of the source: audio/video packets are unchanged, while file-level and stream-level metadata tags, chapters and user-data boxes were removed (leftover box space is a zero-filled free box). Frame counts, timing and resolution are unaffected.

meta.json

{"schema_version": 1, "dataset_version": "v1", "count": N,
 "valid_types": ["body", "left_hand", "right_hand", "face"],
 "target_types": ["pose", "expression", "camera", "identity", "scale"],
 "items": [{"uuid", "private_uuid", "split", "frames", "fps", "width", "height",
            "valid_frames": {"body": n, "left_hand": n, "right_hand": n, "face": n} | null,
            "target_valid_frames": {"pose": n, ...} | null,
            "files": {"video": path, "rgb": path | null}, "bytes": {"video": n, "rgb": n},
            "sha256": {"video": hex, "rgb": hex}}]}

• split: train / val / test.
• frames: number of video frames.
• valid_frames[part]: frames where rgb.npz visible[:, part] is true; null when the clip has no rgb.npz.
• target_valid_frames[type]: frames with at least one valid channel of that 3D body-model target type in the (non-public) training targets; null when the clip has no targets. Counts only.
• private_uuid: identifier of the clip's non-public companion files; nothing under it is in this bucket.