A Conditional Generative Adversarial Network for Non-rigid Point Set Registration

Abstract

This paper proposes a novel approach to perform non-rigid point set registration without an iterative process. The main idea is to design a conditional generative adversarial network, termed Point Registration Generative Adversarial Network (PR-GAN). The proposed PR-GAN establishes an adversarial game between a generator and a discriminator. The generator aims to generate the geometric transformation parameters, and the discriminator aims to force the generated parameters to register two point sets accurately. After effective training, PRGAN can generate the desired transformation parameters to register a never-seen-before point set pair without an iterative optimization process. Furthermore, we design a pre-trained autoencoder to represent the point sets before feeding to PRGAN. Experiments with deformation, noise, and outlier are conducted. Results exhibit that PR-GAN achieves remarkably better performance compared to traditional iterative solutions.