For skeleton-based human action recognition, one of the key challenges is the large view variations when capturing data. In order to alleviate the effects of view variations, we introduce a novel view adaptation scheme, which automatically determines the virtual observation viewpoints in a learning based data driven manner.
This repository holds the codes and methods for the following papers:
View Adaptive Neural Networks for High Performance Skeleton-based Human Action Recognition. TPAMI, 2019.
View Adaptive Recurrent Neural Networks for High Performance Human Action Recognition from Skeleton Data. ICCV, 2017.
Figure 1: Flowchat of the end-to-end view adaptive neural network. It consists of a main classification network and a view adaptation subnetwork. The view adaptation subnetwork automatically determines the virtual observation viewpoints and transforms the skeleton input to representations under the new viewpoints for classification by the main classification network. The entire network is end-toend trained to optimize the classification performance.
Figure 2: Architecture of the proposed view adaptive neural networks: a view adaptive RNN with LSTM (VA-RNN), and a view adaptive CNN (VA-CNN). The classification scores from the two networks can be fused to provide the fused prediction, denoted as the VA-fusion scheme. Note that based on application requirements, we can use VA-RNN or VA-CNN only or combine them together.
Figure 3: Frames of (a) the similar posture captured from different viewpoints for the same subject, and (b) the same action “drinking” captured from different viewpoints for different subjects. 2nd row: original skeletons. 3rd row: Skeleton representations from the observation viewpoints of our VA-RNN model. 4th row: Skeleton representations from the observation viewpoints of our VA-CNN model.
The code is built with following libraries:
We use the NTU60 RGB+D dataset as an example for description. We need to first dowload the NTU-RGB+D dataset
- Extract the dataset to ./data/ntu/nturgb+d_skeletons/
- Process the data
cd ./data/ntu
# Get skeleton of each performer
python get_raw_skes_data.py
# Remove the bad skeleton
python get_raw_denoised_data.py
# Transform the skeleton to the center of the first frame
python seq_transformation.py# For CNN-based model with view adaptation module
python va-cnn.py --model VA --aug 1 --train 1
# For CNN-based model without view adaptation module
python va-cnn.py --model baseline --aug 1 --train 1
# For RNN-based model with view adaptation module
python va-rnn.py --model VA --aug 1 --train 1
# For RNN-based model without view adaptation module
python va-rnn.py --model baseline --aug 1 --train 1# For CNN-based model with view adaptation module
python va-cnn.py --model VA --aug 1 --train 0
# For CNN-based model without view adaptation module
python va-cnn.py --model baseline --aug 1 --train 0
# For RNN-based model with view adaptation module
python va-rnn.py --model VA --aug 1 --train 0
# For RNN-based model without view adaptation module
python va-rnn.py --model baseline --aug 1 --train 0If you find our papers and repo useful, please cite our papers. Thanks!
@article{zhang2019view,
title={View adaptive neural networks for high performance skeleton-based human action recognition},
author={Zhang, Pengfei and Lan, Cuiling and Xing, Junliang and Zeng, Wenjun and Xue, Jianru and Zheng, Nanning},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
year={2019},
}
@inproceedings{zhang2017view,
title={View adaptive recurrent neural networks for high performance human action recognition from skeleton data},
author={Zhang, Pengfei and Lan, Cuiling and Xing, Junliang and Zeng, Wenjun and Xue, Jianru and Zheng, Nanning},
booktitle={Proceedings of the IEEE International Conference on Computer Vision},
pages={2117--2126},
year={2017}
}
Microsoft Open Source Code of Conduct: https://opensource.microsoft.com/codeofconduct