Multi-Task Collaborative Attention Network for Pedestrian Attribute Recognition

Pedestrian Attribute Recognition (PAR) is a multi-task attribute leaning problem. Research into person attributes recognition has focused on approaches to describe a person in terms of their appearance. Combination of some attributes is helpful to strengthen each other's learning such as upper clothing style and upper clothing length, while others are not, such as hair style and upper clothing length. Thus, how to effectively combine different task is the key challenges in PAR. To effectively utilizing the relationship between attributes and further improve the effects of PAR, we propose a novel Multi-Task Collaborative Attention Network (MTCAN), which consists of three modules. Specifically, we first design a Feature Division Module (FDM) to focus on reliable and flexible attribute-related regions. Based on the precise attribute-related locations, we further construct a Spatial and Channel Collaborative Attention Module (SCCAM) to facilitate the beneficial features and weaken mutually suppressed features. Thirdly, a newly weights fusion strategy named adaptive-soups is proposed to mine the optimal model which is universal for deep learning models in all fields. Experiments on two pedestrian attribute recognition datasets show that our proposed method achieves superior performance against other state-of-the-art methods.