Multi-Level Feature Learning for Pedestrian Attribute Recognition

Pedestrian attribute recognition is important for many subjects such as pedestrian tracking and person re-identification in monitoring scenario. Recently plenty of models address this task with deeply learned feature representations, but there still great potentials to make further progress due to some variations including low resolution, occlusion and so on. In this paper, we propose a new deep network structure for attribute classification, which takes advantage of multi-level features and an attention weighted scheme to combine multiple predictions from different layers. At last, we evaluate our method on PA-100K benchmark and the experimental results show the effectiveness of our proposed approach.