3D point cloud domain generalization via adversarial training

Abstract

The purpose of the paper is to tackle the classification problem of 3D point cloud data in domain generalization: how to develop a generalized feature representation for an unseen target domain by utilizing sub-field of numerous seen source domain(s). We present a novel methodology based on both adversarial training to learn a generalized feature representations across subdomains in domain adaptation called 3D-AA. We specifically expand adversarial autoencoders by applying the Maximum Mean Discrepancy (MMD) measure to align the distributions across several subdomains, and then matching the aligned distribution to any given prior distribution via adversarial feature learning. In this manner, the learned 3D feature representation is supposed to be universal to the observed source domains due to the MMD regularization and is expected to generalize well on the target domain due to the addition of the prior distribution. We applied an algorithm to train two different 3D point cloud source domains with our framework. The combination of multiple loss functions on 3D point cloud domain generalization task show that our applied algorithm performs better and learn more generalized features for the target domain than the source-only algorithm which only utilized the MMD measurement.