This package is a PyTorch implementation of our paper 3C-Net: Category Count and Center Loss for Weakly-Supervised Action Localization, to be published at ICCV 2019.
We use the same I3D features, for the Thumos14 and ActivityNet 1.2 datasets, released by Sujoy Paul. The annotations are already included in this repository.
The model can be trained using the following commands. See options.py for additional parse arguments
# Running on Thumos14
python main.py --dataset-name Thumos14
# Running on ActivityNet 1.2
python main.py --dataset-name ActivityNet1.2 --activity-net --num-class 100
Please cite the following work if you use this package.
@article{narayan20193cnet,
title={3C-Net: Category Count and Center Loss for Weakly-Supervised Action Localization},
author={Narayan, Sanath and Cholakkal, Hisham and Shahbaz Khan, Fahad and Shao, Ling},
journal={arXiv preprint arXiv:1908.08216},
year={2019}
}
This codebase was built on the W-TALC package found here and has the following dependencies.
- PyTorch 0.4.1, Tensorboard Logger 0.1.0
- Python 3.6
- numpy, scipy among others