Learning Camera-Invariant Representation for Person Re-identification

Person re-identification (re-ID) problem aims to retrieve a person from an image gallery captured across multiple cameras. However, images of the same identity have variations due to the change in camera views. So learning a camera-invariant representation is one objective of re-identification. In this paper, we propose a camera-style transfer model for generating images, and a fake triplet loss for training the person feature embedding model. We train a StarGAN, a kind of generative adversarial networks, as our transfer model, which can transfer the style of an image from one camera to multiple different camera-styles by a generator network. So the image set is expanded with style-transferred images. However, style transferring yields image distortion, which misleads the training of feature embedding model. To overcome the influence of image distortion, we consider the gap between fake and real images, then we propose a fake triplet loss to capture the camera-invariant information of fake images. We do a series of experiments on the Market-1501, DukeMTMC-reID, and CUHK03 datasets, and show the effectiveness of our methods.