CapsGAN: Using Dynamic Routing for Generative Adversarial Networks
This addresses image generation with geometric transformations for computer vision, though it appears incremental as it combines existing methods.
The paper tackles the problem of generating 3D images from geometrically transformed inputs by combining GANs with capsule networks (CapsGAN), showing it performs better than or equal to CNN-based GANs on rotated MNIST.
In this paper, we propose a novel technique for generating images in the 3D domain from images with high degree of geometrical transformations. By coalescing two popular concurrent methods that have seen rapid ascension to the machine learning zeitgeist in recent years: GANs (Goodfellow et. al.) and Capsule networks (Sabour, Hinton et. al.) - we present: \textbf{CapsGAN}. We show that CapsGAN performs better than or equal to traditional CNN based GANs in generating images with high geometric transformations using rotated MNIST. In the process, we also show the efficacy of using capsules architecture in the GANs domain. Furthermore, we tackle the Gordian Knot in training GANs - the performance control and training stability by experimenting with using Wasserstein distance (gradient clipping, penalty) and Spectral Normalization. The experimental findings of this paper should propel the application of capsules and GANs in the still exciting and nascent domain of 3D image generation, and plausibly video (frame) generation.