Unsupervised face-name association via commute distance

• DOI: 10.1145/2393347.2393383
• Publication Date: 2012-11-14T15:36:17Z
• Authors: Jiajun Bu; Bin Xu; Chenxia Wu; Chun Chen; Jianke Zhu; Deng Cai; Xiaofei He
• Keywords: Adjacency matrix; Association (psychology)

Recently, the task of unsupervised face-name association has received a considerable interests in multimedia and information retrieval communities. It is quite different with generic facial image annotation problem because its ambiguous assignment properties. Specifically, should obey following three constraints: (1) face can only be assigned to name appearing associated caption or null; (2) at most one face; (3) name. Many conventional methods have been proposed tackle this while suffering from some common problems, eg, many them are computational expensive hard make null decision. In paper, we design novel framework named via commute distance (FACD), which judges face-null assignments under unified (CD) algorithm. Then, further speed up on-line processing, propose anchor-based (ACD) algorithm whose main idea using anchor point representation structure accelerate eigen-decomposition adjacency matrix graph. Systematic experiment results on large scale real world image-caption database total 194,046 detected faces 244,725 names show that our approach outperforms state-of-the-art performance. Our appropriate for real-time system.