Spatial Quality Aware Network for Video-Based Person Re-identification

Person re-identification in video is challenging in computer vision. Most methods adopt feature aggregation to get a video-level representation. However, almost all of them do it on the final feature embedding, which neglects the spatial difference among feature maps. To address this problem, we proposed an effective approach, named Spatial Quality Aware Network (SQAN) for video-based person re-identification. SQAN distributes a score for each pixel in a feature map. Then scores are normalized across all frames and the weighted sum is used to aggregate them. To deal with overfitting, we also proposed a semantic dropout strategy. Experiments show that our proposed method is competitive with state-of-the-art methods in performance.