Spatial and Semantic Relations for Pedestrian Attribute Recognition

This paper addresses the problem of pedestrian attribute recognition. Previous works typically treat the different attributes independently with each other, without considering possible dependencies between them, or just take semantic relations or spatial relations into consideration. In our work, we propose an end-to-end learning framework combining with a subnet for multi-task classification to take both spatial and semantic relations into account, which proves to be more accurate and effective. Our work can not only deal with binary attributes, but also multi-class attributes in a single network, overcoming the drawbacks in many methods that can only recognize the binary attributes in joint-training. Experiments have been carried out on several benchmarks and positively demonstrated the superiority of our method.