You are here

ECE PhD Defense: "Low Rank Tensor Learning for Human Action Recognition," Chengcheng Jia


427 Richards Hall

April 18, 2016 12:00 pm
April 18, 2016 12:00 pm


RGB-D action has aroused more attention in action recognition field recently, because the depth information could avoid irrelevant effect of illumination and occlusion etc. Different from face or image recognition which considers 2D data mostly, an action data includes not only spatial information but also temporal knowledge, which is not taken into account in the traditional vectorized methods for visual recognition. Inspired of this, we employ a third-order tensor to represent an action sequence, while the first and second fold (mode) indicate spatial pixels and third fold indicates temporal changing. Considering the high-dimensional property of an action sequence, we use subspace learning method to reduce dimensions for reasonable time complexity. In my dissertation, I mainly introduce three tensor subspace learning models, the first model aims to find the subspace by low-rank learning on the projection matrices, which find the dimension automatically instead of manually defined. The second model transfers the rich information from well-established domain to the existing domain to recover ``missing'' modality, which could improve the performance of single modality. The third model is selecting key frames and aligning them automatically in a subspace, by sparse learning and non-negative tensor factorization.

Advisor: Professor Yun Fu

Professor Yun Fu
Professor Yizhou Sun
Professor Stratis Ioannidis