DIPR: Efficient Point Cloud Registration via Dynamic Iteration

TL;DR

DIPR tackles the computational burden of point cloud registration by dynamically concentrating effort on overlapping regions. It introduces a two-stage global-local framework with Deeper Sampling, Refined Nodes based on density clustering, and an SC Classifier to terminate iterations early, enabling efficient inference with sparser inputs. The method combines a coarse global matching stage with Local-to-Global refinement and subsequent local registration on refined inliers, achieving state-of-the-art Registration Recall (RR) on 3DMatch, 3DLoMatch, and KITTI while reducing runtime and GPU memory. Extensive ablations validate the contributions of two encoders, Refined Nodes, and the iterative scheme. Overall, DIPR offers a practical, scalable solution for fast and accurate PCR in challenging scenarios with varying overlap.

Abstract

Point cloud registration (PCR) is an essential task in 3D vision. Existing methods achieve increasingly higher accuracy. However, a large proportion of non-overlapping points in point cloud registration consume a lot of computational resources while negatively affecting registration accuracy. To overcome this challenge, we introduce a novel Efficient Point Cloud Registration via Dynamic Iteration framework, DIPR, that makes the neural network interactively focus on overlapping points based on sparser input points. We design global and local registration stages to achieve efficient course-tofine processing. Beyond basic matching modules, we propose the Refined Nodes to narrow down the scope of overlapping points by using adopted density-based clustering to significantly reduce the computation amount. And our SC Classifier serves as an early-exit mechanism to terminate the registration process in time according to matching accuracy. Extensive experiments on multiple datasets show that our proposed approach achieves superior registration accuracy while significantly reducing computational time and GPU memory consumption compared to state-of-the-art methods.

Figures (7)

• Figure 1: We propose a dynamic iteration approach to improve the accuracy and efficiency of PCR. Our DIPR samples sparser point clouds and dynamically search for overlapping areas in the local stage by Refined Nodes. After each stage, we design an SC Classifier to determine whether the registration result needs to be continued. Here, $SC$ and $SC{\uparrow}$ are the registration evaluation metrics. “FLOPs” refers to the proportion of the computation required by our method versus RoITR yu2023rotation.
• Figure 1: Registration results on KITTI. Our method achieves higher accuracy after local registration and achieves good performance similar to ground truth.
• Figure 2: Our DIPR is mainly divided into two parts: global registration and local registration. We first design the Deeper Sampling to downsampling and extract the initial feature. Then, we use the Basic Matching Module to obtain the pairing points $\tilde{P}_{X}$ and $\tilde{P}_{Y}$ and the global registration result. Next, the Refined Nodes in the local registration stage clusters the paired points and obtains the new coarse node for local registration. After each registration stage finishes, we design the SC Classifier to determine whether the current result requires further iteration.
• Figure 3: Comparison with existing methods on multiple datasets. Our method achieves SOTA registration accuracy (RR) with much lower time and GPU memory consumption on 3DMatch and 3DLoMatch datasets.
• Figure 4: Visualization results of our network for 3DLoMatch (a-d) and 3DMatch (e-h) with global and local registration. Compared to GeoTrans qin2022geometric, our method has similar results in the global registration, and the local registration is further optimized to achieve superior results.