Activity Recognition in Still Images with Transductive Non-negative Matrix Factorization

摘要

Still image based activity recognition is a challenging problem due to changes in appearance of persons, articulation in poses, cluttered backgrounds, and absence of temporal features. In this paper, we proposed a novel method to recognize activities from still images based on transductive non-negative matrix factorization (TNMF). TNMF clusters the visual descriptors of each human action in the training images into fixed number of groups meanwhile learns to represent the visual descriptor of test image on the concatenated bases. Since TNMF learns these bases on both training images and test image simultaneously, it learns a more discriminative representation than standard NMF based methods. We developed a multiplicative update rule to solve TNMF and proved its convergence. Experimental results on both laboratory and real-world datasets demonstrate that TNMF consistently outperforms NMF.