This repository is the introduction of "PBRNet:Progressive Boundary Refinement Network for Temporal Action Detection"(AAAI2020). They are designed for accurate and efficient temporal action detection. Initially, the framework is implemented by tensorflow and now we re-implement it by pytorch.
python 3.6
pytorch 1.6
opencv-python 3.4.1
You first need to download the THUMOS14 datasets, then extract frames from videos by 10 fps. For optial flow extraction, you can refer to TV-L1 which only requires CPU. You can change the configurations based on your resources.
We refer to kinetics_i3d_pytorch to get the pre-trained i3d model.
CUDA_VISIBLE_DEVICES=$GPU_IDs python main.py
@inproceedings{liu2020progressive,
title={Progressive boundary refinement network for temporal action detection},
author={Liu, Qinying and Wang, Zilei},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2020}
}