Person Re -Identification Using Multi-region Triplet Convolutional Network

Person re-identification is a difficult task due to variations of person pose, scale changes, different illumination, occlusions, to name a few important factors usually diminishing identification performance across different views. In this work, we train a siamese and triplet convolutional neural networks and show that they can achieve promising recognition ratios. In order to cope with spatial transformations and scale changes across multi-view images we employ deformable convolutions in a triplet convolutional neural network. We propose an unified neural network architecture consisting of three triplet convolutional neural networks to jointly learn both the local body-parts features and full-body descriptors. We demonstrate experimentally that it achieves comparable results with results achieved by state-of-the-arts methods.