Abstract

In this work, we investigate how to automatically uncover the underlying group structure of a feature vector such that each group characterizes certain ob ject-specific patterns, e.g., visual pattern or mo- tion tra jectories from one ob ject. By mining the group structure, we can effectively alleviate the mutual inference of multiple ob jects and improve the performance in various visual analysis tasks. To this end, we propose a novel auto-grouped sparse representation (ASR) method. ASR groups semantically correlated feature elements together through optimally fus- ing their multiple sparse representations. Due to the intractability of pri- mal ob jective function, we also propose well-behaved convex relaxation and smooth approximation to guarantee obtaining a global optimal solu- tion effectively. Finally, we apply ASR in two important visual analysis tasks: multi-label image classification and motion segmentation. Com- prehensive experimental evaluations show that ASR is able to achieve superior performance compared with the state-of-the-arts on these two tasks.