Joint Grouping and Labeling via Complete Graph Decomposition

• Published in: Neural Information Processing Vol. 1143; pp. 497 - 505
• Main Authors: Ge, Jinchao, Wang, Zhenhua, Meng, Jiajun, Zhang, Jianhua, Chen, Shengyong
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2019; Springer International Publishing
• Series: Communications in Computer and Information Science
• Subjects: Activity understanding; Graph decomposition; Grouping
• ISBN: 3030368017; 9783030368012
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-3-030-36802-9_53

We introduce the complete graph decomposition approach for joint grouping and labeling. Our framework takes into consideration both how to group subjects and how to assign labels to them in a joint manner, without knowing the number of groups beforehand. We model the relations of different targets via a complete graph, which is decomposed into a set of complete subgraphs to represent distinct groups. We implement this joint framework by fusing both deep features and rich contextual cues with model parameters learned from data. We propose an alternating search algorithm to solve the relevant inference problem efficiently. We evaluate the effectiveness of the proposed approach on human activity understanding, and show the proposed approach is competitive compared against the state-of-the-art.