The paper begins by providing a definition of point cloud registration (PCR) and classifying commonly used datasets and evaluation metrics for PCR tasks.
For supervised PCR algorithms, the paper organizes the techniques into four key stages: descriptor extraction, correspondence search, outlier filtering, and transformation parameter estimation. It also discusses two fundamental concepts: optimization and multimodal. The paper systematically categorizes the supervised algorithms based on their contributions to each stage or integration of these concepts.
For unsupervised PCR algorithms, the paper differentiates between two methodologies: correspondence-free approaches, which align point clouds by minimizing feature discrepancies, and correspondence-based approaches, which align point clouds by establishing correspondences.
The paper highlights open challenges and potential directions for future research in PCR, such as bridging the gap between synthetic and real-world data, exploiting multimodal information, designing new evaluation metrics, and leveraging pre-trained models.
In eine andere Sprache
aus dem Quellinhalt
arxiv.org
Wichtige Erkenntnisse aus
by Yu-Xin Zhang... um arxiv.org 04-23-2024
https://arxiv.org/pdf/2404.13830.pdfTiefere Fragen