Bridging Domain Gap of Point Cloud Representations via Self-Supervised Geometric Augmentation

TL;DR

A novel scheme for induced geometric invariance of point cloud representations across domains is introduced, via regularizing representation learning with two self-supervised geometric augmentation tasks, and pioneer an integration of the self-supervised relational learning on geometrically-augmented point clouds in a cascade manner.

Abstract

Recent progress of semantic point clouds analysis is largely driven by synthetic data (e.g., the ModelNet and the ShapeNet), which are typically complete, well-aligned and noisy free. Therefore, representations of those ideal synthetic point clouds have limited variations in the geometric perspective and can gain good performance on a number of 3D vision tasks such as point cloud classification. In the context of unsupervised domain adaptation (UDA), representation learning designed for synthetic point clouds can hardly capture domain invariant geometric patterns from incomplete and noisy point clouds. To address such a problem, we introduce a novel scheme for induced geometric invariance of point cloud representations across domains, via regularizing representation learning with two self-supervised geometric augmentation tasks. On one hand, a novel pretext task of predicting translation distances of augmented samples is proposed to alleviate centroid shift of point clouds due to occlusion and noises. On the other hand, we pioneer an integration of the relational self-supervised learning on geometrically-augmented point clouds in a cascade manner, utilizing the intrinsic relationship of augmented variants and other samples as extra constraints of cross-domain geometric features. Experiments on the PointDA-10 dataset demonstrate the effectiveness of the proposed method, achieving the state-of-the-art performance.

Figures (6)

• Figure 1: The illustration of resulting t-SNE representation space with and w/o our proposed method for point cloud domain adaptation. The proposed method not only employ translation distance prediction to alleviate centroid shift of point clouds due to occlusion and noises, but also utilize relational learning to further understand the significant topological changes between source and target domains.
• Figure 2: The framework of our proposed method for unsupervised domain adaptation on point clouds. The framework comprises three critical components: translation distance prediction to alleviate centroid shift of point clouds, relational modeling to capture relationships between cross-domain samples, and representation learning through supervised learning to further align representation across domains. These tasks utilize a shared feature encoder, effectively integrating their capabilities to improve the effectiveness of domain adaptation.
• Figure 3: The illustration of self-supervised translation augmentation. The sample is translated along the $x$-axis and $y$-axis, where the translation distance is determined by the maximum span of translation.
• Figure 4: The illustration of relational learning with shape augmentation. Different augmented versions of the same sample are encoded and projected to the feature space, where similarities are calculated with features of other samples in the feature memory bank to derive the corresponding relational distribution. The relational distribution between two pairs are aligned to achieve relationship consistency.
• Figure 5: (a) w/o Adapt: M $\rightarrow$ S*. (b) w/o Adapt: S $\rightarrow$ S*. (c) w/ Adapt: M $\rightarrow$ S*. (d) w/ Adapt: S $\rightarrow$ S*. Confusion matrices for the classification of test samples on the target domain. Darker colors within the visualization reflect higher levels of accuracy.