Improved Occluded Person Re-Identification with Multi-feature Fusion

Pedestrians are often occluded by various obstacles in public places, which is a big challenge for person re-identification. To address this issue, we propose an improved occluded person re-Identification with the feature fusion network. The new network integrates spatial attention and pose estimation (SAPE) to learn representative, robust, and discriminative features. Specifically, the spatial attention mechanism anchors the regions of interest to the unoccluded spatial semantic information. It digs out the visual knowledge that is helpful for recognition from the global structural pattern. Then, we explicitly partition the attention-ware global feature into parts and improve the recognition granularity by matching local features. On this basis, we improve a pose estimation model to extract the information of the key points and feature fusion with the attention-aware feature to eliminate the influence of occlusion on the re-identification result. We test and verify the effectiveness of the SAPE on Occluded-REID, Occluded-DukeMTMC and Partial-REID. The experiment results show that the proposed method has achieved competitive performance to the state-of-the-art.