Self-distillation Enhanced Vertical Wavelet Spatial Attention for Person Re-identification

Yuxuan Zhang, Huibin Tan, Long Lan, Xiao Teng, Jing Ren, Yongjun Zhang

Abstract: Person re-identification is a challenging problem in computer vision, aiming to accurately match and recognize the same individual across different viewpoints and cameras. Due to significant variations in appearance under different scenes, person re-identification requires highly discriminative features. Wavelet features contain richer phase and amplitude information as well as rotational invariance, demonstrating good performance in various visual tasks. However, through our observations and validations, we have found that the vertical component within wavelet features exhibits stronger adaptability and discriminability in person re-identification. It better captures the body contour and detailed information of pedestrians, which is particularly helpful in distinguishing differences among individuals. Based on this observation, we propose a vertical wavelet spatial attention only with the vertical component in the high frequency specifically designed for feature extraction and matching in person re-identification. To enhance spatial semantic consistency and facilitate the transfer of knowledge between different layers of wavelet attention in the neural network, we introduce a self-distillation enhancement method to constrain shallow and deep spatial attention. Experimental results on Market-1501 and DukeMTMC-reID datasets validate the effectiveness of our model.