Unsupervised Template-assisted Point Cloud Shape Correspondence Network

TL;DR

This paper tackles unsupervised point cloud shape correspondence for highly deformable, non-rigid shapes. It introduces TANet, which combines a learnable template bank and an adaptive template assistance pipeline to guide correspondences, employing a template selector, correlation fusion, and a transitive consistency loss. The template generation module builds explicit, structured templates and aligns them to input shapes via a space aligner, while the template assistance module selects the best template and fuses template correlations to stabilize matching. Across benchmarks like TOSCA and SHREC'19, TANet achieves state-of-the-art results and shows strong cross-dataset generalization to SMAL and SURREAL, with robustness to occlusion and noise in real-world scans. These results demonstrate that template-guided, unsupervised correspondences are effective for complex non-rigid objects.

Abstract

Unsupervised point cloud shape correspondence aims to establish point-wise correspondences between source and target point clouds. Existing methods obtain correspondences directly by computing point-wise feature similarity between point clouds. However, non-rigid objects possess strong deformability and unusual shapes, making it a longstanding challenge to directly establish correspondences between point clouds with unconventional shapes. To address this challenge, we propose an unsupervised Template-Assisted point cloud shape correspondence Network, termed TANet, including a template generation module and a template assistance module. The proposed TANet enjoys several merits. Firstly, the template generation module establishes a set of learnable templates with explicit structures. Secondly, we introduce a template assistance module that extensively leverages the generated templates to establish more accurate shape correspondences from multiple perspectives. Extensive experiments on four human and animal datasets demonstrate that TANet achieves favorable performance against state-of-the-art methods.

Figures (11)

• Figure 1: Visualization of template-assisted shape correspondence results. Correspondences are depicted by transferring colors from the source to the target based on matching results. The baseline incorrectly aligns the hands with the head and knees between unconventional shapes. In contrast, our method leverages the template to establish accurate correspondences for the hands.
• Figure 2: Illustration of the TANet. TANet comprises an encoder, a template generation module, and a template assistence module. The template generation module produces several learnable shape templates in the template bank with a space aligner. The template assistance module selects suitable templates for point cloud pairs and improve accuracy via correlation fusion and transitive consistency.
• Figure 3: Space aligner structure. The predicted point positions in the output are constrained to match the true point positions.
• Figure 4: Template selector structure. The best suitable template is determined through geometric and semantic measures of similarity between templates and point cloud pairs.
• Figure 5: The correlation fusion process. Features of point cloud pairs and embeddings of templates compute correlation vectors, which are then fused using an attention mechanism.