Low-Rank Transfer Human Motion Segmentation

This repository contains code for our TIP (Transactions on Image Processing) paper: Low-Rank Transfer Human Motion Segmentation (LTS). LTS is a journal extension on our previous work: Learning Transferable Subspace for Human Motion Segmentation. LTS proposed a novel low-rank constraint associate with the temporal constraint for human action segmentation in time space. Experimental results demonstrates the effectiveness and high performance of the proposed model.

Human motion segmentation has great potential in real world applications. Conventional segmentation approaches cluster data with no guidance from prior knowledge, which could easily cause unpredictable segmentation output and decrease the performance. To this end, we seek to improve the human-motion segmentation performance by fully utilizing pre-existing well-labeled source data. Specifically, we design a new transfer subspace clustering method for motion segmentation with a weighted rank constraint. Specifically, our proposed model obtains representations of both source and target sequences by mitigating their distribution divergence, which allows for more effective knowledge transfer to the target. To guide new representation learning, we designed a novel sequential graph to preserve temporal information residing in both the source and target. Furthermore, a weighted low-rank constraint is added to enforce the graph regularizer and uncover clustering structures within data. Experiments are evaluated on four human motion databases, which prove the enhanced performance and increased stability of our model compared to state-of-the-art baselines.

Running the code

The code is MATLAB code works in Ubuntu system. Windows system may need minor revision in the folder name (e.g., the folder name should revise from xx/xx (Linux version) to xx\xx (Window version)). Directly download the code and run LTS_demo.m, the cold should works well and report the similar performance as reported in our paper. Feel free to send us emails if you have any difficulties running the code.

This repository contains four datasets, Weizmann dataset, Keck dataset, MAD dataset, and UT-Interaction dataset, which are extracted HOG feature frame by frame. The features are 324-dimension feature metrix.

Authors

Welcome to send us Emails if you have any questions about the code and our work :-)

• Lichen Wang Website
• Zhengming Ding Website
• Yun Raymond Fu Website

Citation

This code is corresponding to our TIP 2019 paper below:

@ARTICLE{LTS_Lichen_TIP18, 
  author={Lichen Wang and Zhengming Ding and Yun Fu}, 
  journal={IEEE Transactions on Image Processing}, 
  title={Low-Rank Transfer Human Motion Segmentation}, 
  year={2019}, 
  volume={28}, 
  number={2}, 
  pages={1023-1034},
  doi={10.1109/TIP.2018.2870945},
}

This work is our AAAI 2018 extension work: Learning Transferable Subspace for Human Motion Segmentation. A low-rank constraint is designed to further reconstruct the temporal data structure and improve the performance.

@inproceedings{TSS_Lichen_AAAI18,
  author = {Lichen Wang and Zhengming Ding and Yun Fu},
  title = {Learning Transferable Subspace for Human Motion Segmentation},
  conference = {AAAI Conference on Artificial Intelligence},
  year = {2018},
}

Please cite our paper if you like or use our work for your research, thank you very much!