Human V1 70F: Filtered Human Activity Episodes with Skeleton and Video Data

Human V1 70F is a dataset of 85,856 episodes of human activity, containing 343,424 uploaded files including video and skeleton data. The dataset was created by zywu2115 and last updated on 2026-05-18. It is filtered based on criteria such as a minimum of 70 frames, a minimum visible ratio of 0.3, and a minimum hand motion of 0.005.

Use Cases

• Train human pose estimation models based on skeleton_scenario.mp4 files.
• Develop video captioning models for human activities using caption.pickle files.
• Analyze egocentric human motion patterns based on filtered episodes from sources like EgoDex.
• Benchmark action recognition algorithms on multi-view video data from rgb.mp4 files.

Strengths

• Contains 85,856 episodes, providing a substantial number of activity instances.
• Includes 343,424 uploaded files across multiple modalities (video, skeleton, captions).
• Applies specific quality filters (min_frames: 70, min_visible_ratio: 0.3, min_hand_motion: 0.005) to the data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Aggregated from multiple sources including EgoDex (78,273 episodes), VITRA Ego4D Cooking (16 episodes), VITRA Ego4D Other (13 episodes), and VITRA EPIC (7,554 episodes).

Collection Method

Likely collected from various egocentric and third-person video sources, then filtered and processed.

Freshness

Last updated 2026-05-18 13:20:23; freshness should be verified.