OPEN DATASET

HomeHands-50

Egocentric household activity footage, plus the derived hand pose, segmentation, depth, and narration datasets built on top of it.

Core Data

HomeHands-50

Egocentric footage of everyday household activities captured from a first-person wearable camera. The foundation for all subsequent HomeHands derived datasets.

50 Hours
5.4M Frames
2,700 Videos
1.8M Hands Tracked
7,600 Narrations
8.9M Seg. Masks
25 Tasks
100 Action Classes
RGB Videos
washing a cup
cutting banana
folding clothes
making tea
organising clothes
washing clothes
mopping
sweeping
arranging shoes
Derived Datasets
Hand Tracking
HomeTrace

21-point hand skeleton tracked per frame using MediaPipe HandLandmarker. Wrist, knuckles, and fingertips detected on both hands.

21Keypoints
82%Detect Rate
2Hands
x,y,zCoords
Segmentation
HomeMask

Per-frame instance masks for hands, objects, and surfaces using a fine-tuned segmentation model across all activity clips.

4Masks/Frame
20Classes
PNGFormat
BinaryMask Type
Audio Narration CC
HomeVoice

Live-narrated egocentric recordings with burned-in captions. The person performing each task describes their own actions in real time.

2,700Clips
7,600Sentences
JSONFormat
SRTSubtitles

Want to Contribute Recordings or Pipeline Tools?

We accept new egocentric recordings and pipeline tooling contributions (preprocessing, validation, annotation scripts). Submissions are scored by the same quality-check pipeline used internally — see the Competitions page.

Annotation Quality (50%) Hand visibility, sharpness, and lighting must clear the gates above.
Coverage (50%) Preference for new activities, rooms, or tasks not yet represented.