yuanxiang102e/egoexo-proficiency-cam01

EgoExo Proficiency Assessment Dataset (cam01 view)

A curated subset of Ego-Exo4D for 4-class skill proficiency level recognition from exocentric fixed-camera video (cam01).

Used to fine-tune InternVL3-2B for proficiency assessment, achieving 45.3% accuracy (LoRA, 4-Part Late Fusion).

Dataset Summary

Split	Samples	Novice	Early Expert	Inter. Expert	Late Expert
Train	1,306	6.6%	34.8%	32.1%	26.6%
Val	364	4.7%	34.1%	33.0%	28.3%
Total	1,670

Action Categories

Category	Train	Val
Bouldering	552	154
Basketball	317	79
Dance	298	97
Soccer	61	10
Cooking	45	13
Music	24	7
Other	9	4

Dataset Structure

takes/
└── <take_uid>/
    └── cam01.mp4       # exocentric fixed-camera view, 52-80s @ 30fps, 448px downscaled
data/
├── train.jsonl         # 1306 training annotations
├── train.csv           # same, CSV format for easy preview
├── val.jsonl           # 364 validation annotations
└── val.csv

Annotation Format (JSONL)

Each line:

{
  "take_uid": "uniandes_bouldering_028_7",
  "cam01_video": "takes/uniandes_bouldering_028_7/cam01.mp4",
  "proficiency_level": "Early Expert",
  "action_category": "bouldering"
}

Labels

Label	Description
Novice	Very limited skill, frequent errors, hesitant movements
Early Expert	Basic competency, occasional errors, technique developing
Intermediate Expert	Consistent technique, infrequent errors, good overall control
Late Expert	Expert-level, highly fluent and automatic movements

Usage

Load Annotations

from datasets import load_dataset

ds = load_dataset("your-username/egoexo-proficiency-cam01")
print(ds["train"][0])
# {'take_uid': 'uniandes_bouldering_028_7',
#  'cam01_video': 'takes/uniandes_bouldering_028_7/cam01.mp4',
#  'proficiency_level': 'Early Expert',
#  'action_category': 'bouldering'}

Load with Videos

import json
from pathlib import Path

dataset_root = Path("/path/to/dataset")

with open(dataset_root / "data/train.jsonl") as f:
    samples = [json.loads(l) for l in f]

for s in samples:
    video_path = dataset_root / s["cam01_video"]
    label = s["proficiency_level"]
    # ... your inference code ...

ms-swift JSONL (for fine-tuning with ms-swift)

Convert to ms-swift format using the provided prepare_data.py script from the accompanying code release.

Video Specifications

• Source: Ego-Exo4D, exocentric camera cam01 (fixed viewpoint)
• Resolution: downscaled to 448px
• Duration: 52–80 seconds per clip
• Frame rate: 30 fps
• Sampling used in experiments: 8 frames uniform sampling per video (or per 4-part segment)

Citation

If you use this dataset, please cite the original Ego-Exo4D paper:

@article{grauman2023ego,
  title={Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives},
  author={Grauman, Kristen and Westbury, Andrew and Torresani, Lorenzo and ...},
  journal={arXiv preprint arXiv:2311.13293},
  year={2023}
}

License

This dataset is derived from Ego-Exo4D and follows its CC BY-NC 4.0 license. Non-commercial use only.