关键词: 关键词: 行人重识别, 计算机视觉, 深度学习, 数据增强, 特征鲁棒性

Abstract: Whether the features had better invariance in the significant intra class changes will determine the upper limit of performance of the Person Re-identification model. Environmental light, image resolution change, motion blur and other factors caused color deviation of pedestrian image. These problems led to overfit the color information of the data, thus limiting the performance of the model. By simulating the loss of the color information of the data sample and highlighting the structural information of the sample, it helped the model to learn more robust features. Specifically, during model training, the training batch is randomly selected according to the set probability, and then randomly selected a rectangular area(or the entire image) for each RGB image sample in the selected batch and replaced its pixels with the same rectangular area in the corresponding grayscale image, thus it generated a training image with different grayscale areas. Experimental results on the Person Re-identification baseline demonstrate that the proposed method has helped the model achieve a significant performance improvement of 3.3% on the mAP, and performed well on multiple datasets.

Key words: Keywords: person re-identification, computer vision, deep learning, data augmentation, robustness of features