This repository contains the implementation, models and data (lebel file) for S3D: Stacking Segmental P3D for Action Quality Assessment:
https://ieeexplore.ieee.org/abstract/document/8451364/.
- PyTorch (with python 3.6)
- Keras
- numpy
- scipy
- sklearn
- skvideo
- opencv
- The command to train ED-TCN on ResNet features
python seg_train.py
- The command to train P3D-spaced
python train_diving.py --gpuid=0 --stop=0.79
- The command to train P3D-center on stage 3
python train_diving.py --tcn_range=3 --downsample=2 --gpuid=0 --stop=0.80
- The command to get correlation using SvR/LR (using the extracted P3D features from
./data_files/all_train_v2.npy
and./data_files/all_test_v2.npy
)
python svr.py
The diving videos are from UNLV-Dive dataset. We annotated segmentation labels for this dataset at ./data_files/jump_drop_water_label.txt
Models with weights can be downloaded from google drive.
- checkpoint90.tar is trained on stage 1 (jumping)
- checkpoint91.tar is trained on stage 2 (dropping)
- checkpoint79.tar is trained on stage 3 (entering into water)
- checkpoint92.tar is trained on stage 4 (ending)
- The P3D model (with weights pre-trained on kinetics) is revised from P3D-Pytorch by qijiezhao.
- The ED-TCN model is revised from ED-TCN by colincsl.
@inproceedings{s3d,
title={S3D: Stacking Segmental P3D for Action Quality Assessment},
author={Xiang Xiang and Ye Tian and Austin Reiter and Gregory D. Hager and Trac D. Tran},
booktitle={IEEE ICIP},
year={2018}
}
If there are any questions, please contact Ye Tian [ytian27@jhu.edu] or Xiang Xiang [xxiang@cs.jhu.edu]