Skip to content

Dataset · PerceptronAI

Ego-centric video benchmark for household tasks

PerceptronAI/EgoHome

First-person video benchmark with dense hand action annotations and sub-second segments starting from 100 ms.

What's inside

Contains 137 first-person household video clips with a total duration of about 2 hours (1920x1080, 30 fps) in English. Each video is manually annotated with four types of annotations: hand bounding boxes, subtasks, hand actions, and text descriptions, including ultra-short segments from 100 ms. Data is presented as video, text, and tables.

What you can do with it

  • Evaluating multimodal video models on household action understanding with high temporal precision
  • Training hand detectors and fine-grained manipulation recognition models in first-person videos
  • Testing dense video captioning systems
Tasks
Video classification, Object detection
Languages
English
Size
up to 1K rows
Downloads this month
2.4K
Likes
4

How to load it

from datasets import load_dataset

ds = load_dataset("PerceptronAI/EgoHome")