C2F-TCN: Coarse to Fine Multi-Resolution Temporal Convolutional Network for Temporal Action Segmentation

Official implementation of Coarse to Fine Multi-Resolution Temporal Convolutional Network for Temporal Action Segmentation link

Code for full supervsion version of ‘C2F-TCN: A Framework for Semi- and Fully-Supervised Temporal Action Segmentation’ link published in TPAMI-2023.

Code for semi-supervised version of the same is available at link.

Data download and directory structure:

The I3D features, ground-truth and test split files are similar used to MSTCN++. In the mstcn_data, download additional files, checkpoints and semi-supervised splits can be downloaded from drive . Specifically, this drive link contains all necessary data in required directory structure except breakfast I3D feature files which can be downloaded from MSTCN++ data directory. It also contains the checkpoints files for supervised C2FTCN.

The data directory is arranged in following structure

mstcn_data
- mapping.csv
- dataset_name
- groundTruth
- splits
- results
  - supervised_C2FTCN
    - split1
      - check_pointfile
    - split2

Run Scripts

The various scripts to run the supervised training, evaluation with test augmentation or with test augmentation is provided as example below. Change the dataset_name, to run on a different dataset.

Training C2FTCN for a particular split of a dataset

##### python train.py --dataset_name <gtea/50salads/breakfast> --cudad <cuda_device_number> --base_dir <data_directory_for_dataset> --split <split_number>
Example:
python train.py --dataset_name 50salads --cudad 1 --base_dir ../mstcn_data/50salads/ --split 5

Evaluate C2FTCN without test time augmentation, showing average results from all splits of dataset

##### python eval.py --dataset_name <gtea/50salads/breakfast> --cudad <cuda_device_number> --base_dir <data_directory_for_dataset> --compile_result
Example:
python eval.py --dataset_name 50salads --cudad 2 --base_dir ../mstcn_data/50salads/ --compile_result

Evaluate C2FTCN with test time augmentation, showing average results from all splits of dataset

##### python eval.py --dataset_name <gtea/50salads/breakfast> --cudad <cuda_device_number> --base_dir <data_directory_for_dataset>
Example:
python eval.py --dataset_name 50salads --cudad 2 --base_dir ../mstcn_data/50salads/

Citation:

If you use the code, please cite

D. Singhania, R. Rahaman and A. Yao, "C2F-TCN: A Framework for Semi- and Fully-Supervised Temporal Action Segmentation," in IEEE Transactions on Pattern Analysis and Machine Intelligence, doi: 10.1109/TPAMI.2023.3284080.

Singhania, D., Rahaman, R., & Yao, A. (2022, June). Iterative contrast-classify for semi-supervised temporal action segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 36, No. 2, pp. 2262-2270).

dipika-singhania/C2F-TCN