InstrAct

Official project repository for “InstrAct: Towards Action-Centric Understanding in Instructional Videos” (ECCV 2026 Poster).

• Project page
• Model training and evaluation
• Data-curation pipeline — Main code and prompts released; detailed scripts and usage instructions will be updated later
• InstrAct Bench annotations
• Paper

Repository structure

InstrAct/
├── docs/             # GitHub Pages project website
├── model/            # InstrAct model, training, and evaluation code
└── data_curation/    # LLM-assisted data-curation code and prompts

Please follow model/README.md for environment setup, backbone checkpoints, annotation formats, training, and evaluation.

Citation

@inproceedings{yang2026instract,
  title     = {InstrAct: Towards Action-Centric Understanding in Instructional Videos},
  author    = {Yang, Zhuoyi and Yu, Jiapeng and Tan, Reuben and Li, Boyang Albert and Xu, Huijuan},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

License

Third-party components retain their original licenses and attribution. A license for the original InstrAct code will be added before the public code release.