KienNguyen1233/EgoExoLearn-processed

EgoExoLearn — processed

Processed derivatives of EgoExoLearn (Huang et al., CVPR 2024) prepared for research on egocentric hand-gesture video generation.

The raw videos are not included. Download them from the original release: hyf015/EgoExoLearn.

Contents

Path	Description
clean_data/	Cleaned ego segments (segment_same/, segment_cross/), 224px exo crops, exo labels, label_same.csv / label_cross.csv, and a PyTorch dataset.py
latent_cache/wan21vae_832x480_f81_p21/	Precomputed Wan2.1-VAE latents (832×480, 81 frames, 21 prefix frames, fp16) as .safetensors. index*.csv maps each sample to its file; meta.json lists the settings
index/	Segment index, ego↔exo alignment and matching, LLM rerank outputs, prototype samples, and splits (splits_prototype.json)
labels_long.csv, labels_downloaded.csv	Narration labels with hand-centric prompt variants
ego_exo_pairs.csv	Ego/exo video pairings
restructure/, build_*.py	Processing scripts that produced the files above
repo/	Snapshot of the official EgoExoLearn code repo (MIT)

Note: to stay under the Hub's 10k-files-per-folder limit, latent_cache/.../same/train/ is split into subfolders named by the first character (lowercased) of each filename, for example same/train/0/00bjmTkN_c0.safetensors. The file column in index*.csv already uses these paths.

Download

hf download KienNguyen1233/EgoExoLearn-processed --repo-type dataset --local-dir EgoExoLearn

License & attribution

Released under the MIT license, the same as the original EgoExoLearn. All credit for the underlying data goes to the EgoExoLearn authors (OpenGVLab, Shanghai AI Lab).

@inproceedings{huang2024egoexolearn,
  title     = {EgoExoLearn: A Dataset for Bridging Asynchronous Ego- and Exo-centric View of Procedural Activities in Real World},
  author    = {Huang, Yifei and Chen, Guo and Xu, Jilan and Zhang, Mingfang and Yang, Lijin and Pei, Baoqi and Zhang, Hongjie and Dong, Lu and Wang, Yali and Wang, Limin and Qiao, Yu},
  booktitle = {CVPR},
  year      = {2024}
}