Variance Weight Distribution Network Based Noise Sample Learning for Robust Person Re-identification

Xiaoyi Long, Ruimin Hu, Xin Xu

Abstract: Person re-identification (re-ID) usually requires a large amount of well-labeled training data to learn generalized discriminative person feature representations. Most of current deep learning models assume that all training data are correctly labeled. However, noisy data commonly exists due to incorrect labeling and person detector errors or occlusions in large scale practical applications. Both types of noisy data can influence model training, while they are ignored by most re-ID models so far. In this paper, we propose a robust deep re-ID model, called variance weight distribution network (VWD-Net), to address this problem. Different from the traditional representations of each person image as a feature vector, the variance weight distribution network focuses on the following three aspects. 1) An improved Gaussian distribution and its variance are used to represent the uncertainty of person features. 2) A well-designed loss in the variance weight distribution network is used to delegate the distribution uncertainty with respect to the training data. 3) The noisy labels are rectified for further optimization on the model training performance. The large scale variance/uncertainty has been assigned to noisy samples and then rectifies their labels, in order to mitigate their negative impact on the training process. Extensive experiments on two benchmarks demonstrate the robustness and effectiveness of VWD-Net.