Efficient Two-Step Networks for Temporal Action Segmentation
This repository provides a PyTorch implementation of the paper Efficient Two-Step Networks for Temporal Action Segmentation.
Requirements
* Python 3.8.5
* pyTorch 1.8.1
You can download packages using requirements.txt.
pip install -r requirements.txt
Datasets
- Download the data provided by MS-TCN, which contains the I3D features (w/o fine-tune) and the ground truth labels for 3 datasets. (~30GB)
- Extract it so that you have the
data
folder in the same directory astrain.py
.
directory structure
├── config
│ ├── 50salads
│ ├── breakfast
│ └── gtea
├── csv
│ ├── 50salads
│ ├── breakfast
│ └── gtea
├─ dataset ─── 50salads/...
│ ├─ breakfast/...
│ └─ gtea ─── features/
│ ├─ groundTruth/
│ ├─ splits/
│ └─ mapping.txt
├── libs
├── result
├── utils
├── requirements.txt
├── train.py
├── eval.py
└── README.md
Training and Testing of ETSN
Setting
First, convert ground truth files into numpy array.
python utils/generate_gt_array.py ./dataset
Then, please run the below script to generate csv files for data laoder'.
python utils/builda_dataset.py ./dataset
Training
You can train a model by changing the settings of the configuration file.
python train.py ./config/xxx/xxx/config.yaml
Evaluation
You can evaluate the performance of result after running.
python eval.py ./result/xxx/xxx/config.yaml test
We also provide trained ETSN model in Google Drive. Extract it so that you have the result
folder in the same directory as train.py
.
average cross validation results
python utils/average_cv_results.py [result_dir]
Citation
If you find our code useful, please cite our paper.
@article{LI2021373,
author = {Yunheng Li and Zhuben Dong and Kaiyuan Liu and Lin Feng and Lianyu Hu and Jie Zhu and Li Xu and Yuhan wang and Shenglan Liu},
journal = {Neurocomputing},
title = {Efficient Two-Step Networks for Temporal Action Segmentation},
year = {2021},
volume = {454},
pages = {373-381},
issn = {0925-2312},
doi = {https://doi.org/10.1016/j.neucom.2021.04.121},
url = {https://www.sciencedirect.com/science/article/pii/S0925231221006998},
}
Contact
For any question, please raise an issue or contact.
Acknowledgement
We appreciate MS-TCN for extracted I3D feature, backbone network and evaluation code.
Appreciating Yuchi Ishikawa shares the re-implementation of MS-TCN with pytorch.