Unsupervised Domain Mapping: Geometry-Consistent GAN

This lecture discusses unsupervised domain mapping using a geometry-consistent generative adversarial network (GCGAN). The instructor explains the concept of cycle consistency and distance preservation in domain mapping, highlighting the limitations and proposing a new approach based on geometry-consistency constraints. The GCGAN model is presented, which takes original images and their geometrically transformed counterparts to enable one-sided unsupervised domain mapping. The lecture compares GCGAN with baseline methods like GAN alone, CycleGAN, and DistanceGAN, showcasing its effectiveness through quantitative and qualitative evaluations on various image translation tasks. The potential applications of GCGAN in producing realistic images are demonstrated through examples like Horse → Zebra, Monet → Photo, and Day → Night transformations.