This paper describes a GAN approach to addressing a common and important problem in person re-identification: inter- and intra-view pose variation. The technique, in extreme synthesis, uses a generative model to implicitly marginalize away pose- and background-dependent information in the feature representaiton to distill a representation that is invariant to both, but still discriminative for person identities. Pose is represented as the spatial configuration of landmarks, and during training person images conditioned on a randomly selected pose are generated from image encodings. These generated images are fed to multiple adversarial discriminators that determine if the generated image is real/false, if the pose in a real/fake image is accurate, and if two feature embeddings correspond to the same person. Experimental results are given on multiple, important benchmark datasets and show significant improvement over the state-of-the-art. Clarity, quality, and reproducibility: The clarity of exposition is quite good.

Person re-identification (reID) is an important task that requires to retrieve a person's images from an image dataset, given one image of the person of interest. For learning robust person features, the pose variation of person images is one of the key challenges. Existing works targeting the problem either perform human alignment, or learn human-region-based representations. Extra pose information and computational cost is generally required for inference. To solve this issue, a Feature Distilling Generative Adversarial Network (FD-GAN) is proposed for learning identity-related and pose-unrelated representations. It is a novel framework based on a Siamese structure with multiple novel discriminators on human poses and identities.