A sparse graph wavelet convolution neural network for video-based person re-identification

Highlights：

• We propose an adaptive graph generation module to capture the semantic relationship between different patches across frames to solve the short time occlusion and pedestrian misalignment problems in videobased person Re-ID.

• The generated sparse weighted graph only captures the supplementary information between highly correlated patches, avoiding the use of pairwise feature mapping that produces information redundancy.

• We propose a sparse graph wavelet convolution neural network (SGWCNN) framework to model and propagate spatial-temporal relationships between different patches across frames to generate more robust and discriminative features.

• Compared with traditional GCN-based methods, SGWCNN has sparse parameters, efficient algorithm, and higher accuracy over large-scale video-based person Re-ID datasets.