Triplet erasing-based data augmentation for person re-identification

Abstract

Occlusion is a fundamental but challenging problem in person re-identification. Previous work like random erasing randomly selects a rectangle region in an image and erases its pixels with random values without considering the correlation between images when triplet loss is employed. To address the problem, we propose an end-to-end approach called triplet erasing-based data augmentation for person re-identification (ReID). We apply this approach to train a convolutional neural network (CNN) with two branches. Local distance branch determines the location of the part that needs to be erased in the image, and then triplet erasing branch erases a rectangle region in the determined part. By generating a variety of occlusion samples, triplet erasing improves the robustness of the model against occlusion. Triplet erasing can increase the distance between the positive sample pairs and decrease the distance between the negative sample pairs, thereby improving the generalisation ability of the network.