CVPR22-POC

This is the official code for our CVPR 2022 paper "Set-Supervised Action Learning in Procedural Task Videos via Pairwise Order Consistency".

Dataset

• Please create a dataset folder and download datasets from this shared folder.

• The features for Breakfast are stored in Breakfast features folder. Please download and rename it as dataset/Breakfast/features.

• The action_occurrence.txt in each dataset folder is created by src/utils/generate_action_occurrence_txt.py.

Training

To train model on a dataset, e.g., Breakfast, please use command:

python -m src.train \
    --cfg src/configs/breakfast.yaml \
    --runs 1 2 3 \
    --set dataset.split 1 gpu 7

It will train POC model for three runs. Network checkpoints and evaluation results will be saved to log folder.

More training options are included in src/configs/default.py.

Citation

Please cite us with

@InProceedings{Lu_2022_CVPR,
    author    = {Lu, Zijia and Elhamifar, Ehsan},
    title     = {Set-Supervised Action Learning in Procedural Task Videos via Pairwise Order Consistency},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {19903-19913}
}