Unsupervised cross modality person re-identification

Abstract

Unsupervised Person Re-Identification (Re-ID) suffers severely from the gap in the modality. Many factors pose a challenge to the task, including occlusions, lightning conditions, pose changes, among several others. Various works try to use different methods to address the issue while we tried to solve it using GANs. We created images of another domain, conserving the identity of the person while changing the modality. It may so happen that a person moves from a well-lit area to an area where the light is way too low to be detected by the visual sensors. In such a case, the camera switches to IR, and the camera gets images in the Infrared spectrum. The method adopted for the generation of images is cycleGAN combined with pose loss and identity loss which further comprises of style loss and content loss. We are intent on getting IR images of a person with different pose whose RGB photos we have while preserving the identity. Besides, we aim to apply the existing state of the art techniques for Unsupervised Person Re-Identification for gauging our images.