Memory-Efficient Point Cloud Registration via Overlapping Region Sampling

TL;DR

This work proposes an overlapping region sampling method, using a k-nearest-neighbor (kNN) based point compression mechanism with multi layer perceptron (MLP) and transformer architectures, that enables efficient large-scale point cloud registration in resource-constrained environments.

Abstract

Recent advances in deep learning have improved 3D point cloud registration but increased graphics processing unit (GPU) memory usage, often requiring preliminary sampling that reduces accuracy. We propose an overlapping region sampling method to reduce memory usage while maintaining accuracy. Our approach estimates the overlapping region and intensively samples from it, using a k-nearest-neighbor (kNN) based point compression mechanism with multi layer perceptron (MLP) and transformer architectures. Evaluations on 3DMatch and 3DLoMatch datasets show our method outperforms other sampling methods in registration recall, especially at lower GPU memory levels. For 3DMatch, we achieve 94% recall with 33% reduced memory usage, with greater advantages in 3DLoMatch. Our method enables efficient large-scale point cloud registration in resource-constrained environments, maintaining high accuracy while significantly reducing memory requirements.

Figures (3)

• Figure 1: Comparison of point cloud registration using random sampling versus our overlapping region sampling method. Top: original source and target clouds with ground truth alignment. Red dashed boxes: true overlapping regions. Both methods reduce points to 20% of original count. Middle: random sampling uniformly reduces points, potentially missing crucial overlapping areas. Bottom: our method concentrates points in overlapping regions (red dashed boxes) while reducing non-overlapping areas. Our sampling leads to more accurate registration, closely matching ground truth, despite using same number of sampled points.
• Figure 2: Overview of our proposed method to sample from overlap region for efficient 3D point cloud registration. The process consists of three main stages: (1) Point Compression, which includes point reduction, kNN-based feature extraction, and cross attention between source and target point clouds; (2) Overlap Region Estimation using the PREDATOR framework; and (3) Sampling, which involves score propagation and overlap-aware point selection. This approach enables memory-efficient processing while preserving crucial geometric information for accurate registration.
• Figure 3: Comparison of registration recall versus memory usage for different point cloud sampling methods, all using GeoTransformer for registration, on the 3DMatch and 3DLoMatch datasets. The graph compares various sampling techniques including random sampling (RS), Poisson disk sampling (PDS), voxel grid sampling (VGS), and farthest point sampling (FPS), as well as the proposed method with and without points compression (PC). Solid lines represent performance on 3DMatch, while dashed lines show results for the more challenging 3DLoMatch dataset. The proposed sampling method ('Ours') consistently achieves higher recall rates with lower memory usage across both datasets, particularly when using point compression.