Regression Before Classification for Temporal Action Detection

This repository contains the code for the ICASSP2020 paper "Regression Before Classification for Temporal Action Detection".

Highly Recommended:

Download preprocessed data by visiting Onedrive provided by Decouple-SSAD.

Environment

• TensorFlow
• One or more GPU with 12G memory
• pandas, numpy

Have been tested on Ubuntu 16.04

• CUDA 9.0, Python3.5/3.6, tensorflow 1.12.0

Run code

1. (Optinal) Prepare THUMOS14 data.
2. Specify your path in config.py (e.g. feature_path, get_models_dir(), get_predict_result_path()).
3. Modify run.sh to specify gpu(s) device and other parameters according to your need.

Training logs are saved at logs folder. Tensorboard logs are saved at logs folder too. Models are save at models folder. Results are saved at results folder.

Citation

If you like this paper or code, please cite us:

@INPROCEEDINGS{9053319,
  author={C. {Jin} and T. {Zhang} and W. {Kong} and T. {Li} and G. {Li}},
  booktitle={ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Regression Before Classification for Temporal Action Detection}, 
  year={2020},
  pages={1-5},
  doi={10.1109/ICASSP40776.2020.9053319}}

Reference

This implementation largely borrows from Decouple-SSAD by Yupan Huang.