agow1/ego-masquerade

Hand-Object Interaction Dataset

Egocentric hand-object interaction dataset with HOI detections in EPIC-Kitchens format.

Dataset Summary

Task	Videos
assembling	11
group_objects	10
pick_n_place	20
put_bin	10
Total	51

All frames are extracted at 1920x1080 resolution.

Directory Structure

<task_name>/
  <video_id>/
    rgb_frames.zip    # Zipped extracted frames (frame_0000001.jpg, ...)
    hand_det.pkl      # Phantom hand detection format

File Descriptions

• rgb_frames.zip: Zipped directory of extracted video frames at 1920x1080 resolution. Frame filenames follow the pattern frame_NNNNNNN.jpg.
• hand_det.pkl: Pickled hand detections in Phantom format. Contains per-frame bounding boxes for hands (left/right) and active objects with contact state.

Usage

import pickle
import zipfile
from pathlib import Path

# Extract frames
with zipfile.ZipFile("assembling/1/rgb_frames.zip", "r") as z:
    z.extractall("assembling/1/rgb_frames")

# Load hand detections
with open("assembling/1/hand_det.pkl", "rb") as f:
    hand_dets = pickle.load(f)

Processing Pipeline

1. Frame extraction — Extract frames from source videos at 1920x1080
2. Hand-object detection — Faster R-CNN based detector to extract hand and object bounding boxes
3. Detection conversion — Convert raw detections to EPIC-Kitchens format
4. Phantom conversion — Convert to Phantom hand detection format

Citation

If you use this dataset, please cite:

@misc{agow-hoa-dataset,
  title={Hand-Object Interaction Dataset},
  author={AGOW},
  year={2026},
}