Core Concepts
ReLaTo introduces a novel approach for large transformation point cloud registration, outperforming existing methods.
Abstract
The article discusses the challenges of point cloud registration, introduces the ReLaTo architecture, and compares its performance with existing methods. It covers the importance of unsupervised point correspondence, the effectiveness of coarse and refined registration steps, and the impact of different pooling values on the network's performance.
Introduction
Point cloud registration is crucial in various applications.
Existing methods focus on small transformations, but large transformations are common in practical scenarios.
Proposed Approach
ReLaTo architecture for large transformations.
Softmax pooling for confidence scores and matching.
Target-guided denoising for refinement.
Experimental Setup
Evaluation on ModelNet40 and KITTI datasets.
Training parameters and metrics used for assessment.
Results
Comparison with state-of-the-art methods on ModelNet40 and KITTI datasets.
Analysis of unsupervised point correspondence and the network's performance.
Evaluation of coarse and refined registration steps.
Impact of different pooling values on the network's performance.
Stats
"The network is capable of finding correct correspondences between the source and target points even when there are large rotations."
"The refinement step reduces the translation error by between 0.2 and 0.1 meters on every rotation level."
Quotes
"The proposed method successfully registers the point sets, even under large transformations, where other current methods fail most of the time."