Feature Fusion with Non-local for Pedestrian Attribute Recognition

Abstract

In the Pedestrian Attribute Recognition (PAR) research topic, how to extract comprehensive features to represent pedestrian attributes is still an open problem due to its multi-label nature. In order to tackle this problem, we proposed a new end-to-end PAR network based on Feature Fusion with a Non-local operation(FFNL), named FFNL-net. Compared with existing feature fusion approaches that only mechanically pays attention to feature maps from multiple levels, it also exploits the strongest semantic but still higher resolution feature maps. Firstly, the pedestrian image is extracted from a backbone network. Then, three feature maps from different levels and scales are obtained, and the Non-local operation on the above multiple feature maps are applied respectively to further extract spatial information. Finally, a novel feature fusion strategy is performed on them to fetch a fused feature map. Our experimental results on existing popular pedestrian attribute datasets of PETA, PA-100K, and RAP prove that our proposed approach achieves state-of-the-art results.