TGE: Trainable Graph Encoding for point cloud registration

doi:10.1016/j.image.2026.117604

TGE: Trainable Graph Encoding for point cloud registration

2026-09-29

0

PRE

AI

DOI：10.1016/j.image.2026.117604

Original

Original request for help

Save

Abstract

En 中文

With the emergence of Industry 4.0, point cloud registration has become a pivotal element in numerous applications including collaborative defect detection and the assembly of digital twins, and so on. This paper proposes trainable graph encoding (TGE), a comprehensive registration model that leverages the inherent spatial structure of point cloud data. Traditional methodologies in point cloud registration suffers from limitations such as small convergence range, insufficient for managing low overlap scenarios, loss of inherent geometric information and diminished retention of complex 3D structures. In contrast, TGE addresses rigid rotations based on inferred correspondences, thereby offering an efficient point-wise strategy. TGE represents the point cloud as a three-dimensional graph, with nodes capturing both spatial coordinates and centrality within the structure. It has an excellent performance in robust key point matching by utilizing a three-dimensional graph structural information enhanced Transformer. In particular, superpoint features were incorporated as learnable embeddings into the attention layer, thereby encoding more detailed geometric prior information. Moreover, the architectural design of TGE enables the potential for future extension to encompass non-rigid registration tasks. Extensive evaluations have demonstrated that TGE is a dominant approach across a range of datasets. Notably, it achieves significant improvements in registration recall compared to baselines, with gains of 2% and 4.3% on the 3DMatch and 3DLoMatch datasets, respectively. These results serve to reinforce the efficacy of TGE as a robust tool for the detection of defects and the implementation of other 3D registration applications.

Keywords:

Point cloud registration

Graph encoding

Transformer

Defect detection

TGE: Trainable Graph Encoding for point cloud registration

Abstract

Journal

Organization

Cited Papers