Résumé: Person search, which aims to find the target person in the whole scene images, has become a crucial application in security surveillance with various types of noise. Existing methods solve this problem by addressing two subtasks simultaneously: detection, distinguishing people from background noise; and identification, separating target from others noise. However, these two tasks are fundamentally contradictory, as detection is designed to be robust to intra-class variations to capture general human features, while identification is designed to be sensitive to inter-class variations to focus only on the target. Inspired by the observation that the police first narrowed the scope to a small group of people similar to the suspect, and then located the target, we propose a Dual-Focus framework consisting of Coarse-Grained Focus (search for similar people) and Fine-Grained Focus (find the target). Extensive experiments on datasets such as CUHK-SYSU and PRW are conducted to evaluate the effectiveness of our proposed methods. The evaluations demonstrate that our method delivers superior results compared with the state-of-art methods.

Langue: Anglais