From the source
Perceptron is releasing and open-sourcing EgoHome, a benchmark of densely human-annotated egocentric video, alongside its Perceptron Egocentric 1.0 announcement.
- The benchmark contains 137 clips of people doing household tasks, two hours in total, and is the first to combine four annotation types on every video.
- Those types are hand location bounding boxes, subtask annotations, left and right hand actions drawn from Ego4D's closed-vocab schema, and left and right hand captions.
- Perceptron says the release includes raw annotations and code for evaluation metrics, including a new LLM-as-judge measure it calls permissible subsegment Jaccard.
- That metric allows ground truth and predicted segments to match many-to-many, because annotators may divide the same video into one task or many smaller segments.
- The annotations are temporally fine-grained: 3,371 segments under 1 s and 773 under 0.5 s.
- Hand boxes are annotated at 10 fps and smoothed up to 30 fps, and the Ego4D action schema offers 118 actions, of which 47 occur in the release.






