3D point cloud analytics

Summarise with VIRAT
Authors: Sheng Da Ong
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Geometric constraints between feature matches are critical in 3D point cloud registration problems, particularly in scenarios characterised by low spatial overlap and significant sensor noise. Existing approaches typically model unordered matches as a consistency graph, and sample consistent matches to generate transformation hypotheses. However, graph construction often introduces noise and topological redundancy, posing great challenges for handcrafted geometric constraints to render reliable consistency in geometrically degenerate environments, such as those dominated by planar surfaces. This research investigates whether an augmented self-attention mechanism within a Transformer architecture can autonomously learn high-order topological relationships to prune outliers more effectively than traditional rigid graphs. To address these challenges, this paper proposes TransformerGeoPrunerCA, a geometry-aware Transformer encoder designed for robust correspondence pruning. The architecture incorporates a Second Order Graph (SOG) compatibility matrix directly into the attention mechanism to provide a strong geometric inductive bias based on distance preservation. Furthermore, this encoder utilises Spatial-Channel Attention, adapting the channel-wise context fusion concepts from the LeCoT framework to 3D point cloud analytics to enhance the discriminative power of learned 128-dimensional features. The inference pipeline is further strengthened by a robust sampler utilising Spatial Non-Maximum Suppression (NMS) for seed diversification and Power Iteration for local consensus verification. The proposed framework was rigorously evaluated on the 3DLoMatch benchmark, which restricts mutual overlap to 10%-30%. Experimental results demonstrate that the TransformerGeoPrunerCA architecture achieved significant improvement when compared to the vanilla Transformer baseline. While a performance gap remains when compared to hypergraph-based state-of-the-art models like HyperGCT, this paper confirms that integrating geometric priors and spatial-channel attention effectively guides latent topological learning in low-overlap environments.

Fields of Study & Subjects

3D Shape Modeling and Analysis Advanced Image and Video Retrieval Techniques 3D Surveying and Cultural Heritage

Publication details

Journal / source
DR-NTU (Nanyang Technological University)
Publication date
1 January 2026
Language
EN