Generative Adversarial Networks method of polarization 3D reconstruction based on meta-transfer learning

This paper makes a first attempt to combines meta-transfer learning and generative adversarial networks (GAN) technology to solve the Shape from Polarization (SfP) problem. To solve this physics-based ill-posed problem, some researchers choose to blend these physical models as priors into a neural network architecture however cannot meet the requirements of cross-domain and few-shot polarization data. This proposed approach put forward two innovative points. First, we design the meta-transfer method so adapt GAN for few-shot learning tasks. Second, we introduce physical priors between monocular polarization sequence and 3D normal vector into the generative loss term. We report to exceed the previous state-of-the-art on deepsfp dataset, showing the potential of meta-transfer learning in few-shot generative tasks.