Datasets:

macrodata
/

egocentric-hand-benchmark

pretty_name: Egocentric Hand Benchmark Annotations
license: other
configs:
  - config_name: egodaily
    data_files: data/egodaily/*.parquet
  - config_name: hot3d
    data_files: data/hot3d/*.parquet

Egocentric Hand Benchmark Annotations

Redistribution-safe metadata for an egocentric hand benchmark. This repository contains no source RGB, videos, or rendered hand crops. Obtain the RGB from the corresponding source releases under their own terms.

Subsets

`egodaily`

95,002 official manual L/R hand annotations. Each row represents one wearer hand in native 1920×1080 GoPro SuperView image coordinates. Boxes cover the visible hand through the wrist. Use the official EgoDaily image archive and source_image_path to resolve each frame.

Source: Cruz & Chan, 2019.

`hot3d`

5,425 released HOT3D virtual hand-crop camera records from the local HOT3D-Clips subset. These are not 2D detection boxes. Each row identifies a wearer side, source stream, and crop-camera transform/FOV needed to reproduce the released crop rendering after obtaining HOT3D-Clips.

Source: HOT3D.

Schema conventions

side is anatomical left or right.
EgoDaily owner is always wearer.
EgoDaily boxes are bbox_x, bbox_y, bbox_width, bbox_height in source image pixels.
HOT3D is intentionally a separate subset because its crop-camera geometry is not comparable to manual visible-pixel detection boxes.