WhissleAI/egocentric-activity-sample

Egocentric Activity Sample Dataset

A small-scale egocentric (first-person) video dataset with Ego4D-style annotations, designed for quick prototyping and experimentation with egocentric video understanding tasks.

Dataset Summary

Metric	Value
Video clips	19
Total duration	~9.5 minutes
Resolution	960x540 (540p)
FPS	30
Narrations	99
NLQ queries	57
Moment annotations	19
FHO actions	57
Total size	~54 MB

Activities Covered

Scenario	Activities	Clips
Object Manipulation	Pick & place, reorient & place, bimanual manipulation	10
Cleaning	Organizing bathroom, tidying bedroom	6
Cooking	Washing dishes at the sink	3

Dataset Structure

├── videos/                    # 19 egocentric video clips (30s each, 540p, h264)
│   ├── ego_pick_place_01.mp4
│   ├── ego_washing_dishes_01.mp4
│   └── ...
├── annotations/
│   ├── metadata.json          # Video metadata (UIDs, duration, resolution, scenarios)
│   ├── narrations.json        # Dense temporal narrations (Ego4D narration format)
│   ├── nlq.json               # Natural Language Queries (Ego4D NLQ format)
│   ├── moments.json           # Temporal activity moments (Ego4D moments format)
│   ├── fho_actions.json       # Forecasting Hands & Objects actions (Ego4D FHO format)
│   └── taxonomy.json          # Activity/verb/noun taxonomy
└── metadata.csv               # Flat CSV for HuggingFace datasets library

Annotation Formats

All annotations follow the Ego4D v2 annotation schema.

Narrations

Dense temporal narrations using #C (camera wearer) and #O (other person) tags:

{
  "timestamp_sec": 8.0,
  "narration_text": "#C C applies soap to the sponge",
  "is_camera_wearer": true
}

Natural Language Queries (NLQ)

Temporal grounding queries with response windows:

{
  "query": "What dish did I wash?",
  "clip_start_sec": 1.0,
  "clip_end_sec": 24.0
}

Moments

Temporal activity localization labels:

{
  "label": "wash_dishes_/_utensils_/_bakeware_etc.",
  "start_time": 1.0,
  "end_time": 28.0
}

FHO Actions

Hands & objects interaction annotations with critical frames:

{
  "structured_verb": "scrub",
  "structured_noun": "dish",
  "critical_frames": {
    "pre_frame": {"sec": 11.5},
    "contact_frame": {"sec": 13.0},
    "pnr_frame": {"sec": 12.7},
    "post_frame": {"sec": 14.5}
  }
}

Usage

With HuggingFace Datasets

from datasets import load_dataset
ds = load_dataset("WhissleAI/egocentric-activity-sample")

Direct JSON Loading

import json
with open("annotations/narrations.json") as f:
    narrations = json.load(f)
for video in narrations["videos"]:
    for n in video["narrations"]:
        print(f"[{n['timestamp_sec']:.1f}s] {n['narration_text']}")

Sources

Video clips are derived from publicly available egocentric video datasets:

• HoyerChou/EgocentricVideos — pick-place, reorient, bimanual manipulation
• TrainThemAI/POV-Egocentric-Video-Robotics-FHD-Samples — household activities

All videos downscaled to 540p and trimmed to 30-second clips.

License

CC-BY-4.0 — see source datasets for their respective licenses.

Citation

If you use this dataset, please cite the original Ego4D paper for the annotation format:

@inproceedings{grauman2022ego4d,
  title={Ego4d: Around the world in 3,000 hours of egocentric video},
  author={Grauman, Kristen and others},
  booktitle={CVPR},
  year={2022}
}