Primitive fitting based on the efficient multiBaySAC algorithm.

Area of Science:

• Computer Vision
• Geometric Modeling
• Point Cloud Processing

Background:

• Random Sample Consensus (RANSAC) is robust but computationally expensive due to random hypothesis selection and outlier contamination.
• Fitting multiple geometric primitives in unorganized 3D point clouds is a challenging task with existing methods suffering from segmentation issues.

Purpose of the Study:

• To develop a novel algorithm, multiBaySAC (Bayes Sample Consensus), for efficient and accurate fitting of multiple primitives in 3D point clouds.
• To improve upon the computational efficiency and accuracy of traditional RANSAC methods in multiple primitive fitting scenarios.

Main Methods:

• Introduced a statistical testing algorithm for candidate model parameter histograms to detect initial primitives.
• Implemented a parallel conditional sampling method, BaySAC, assigning multiple prior inlier probabilities to each data point for initial primitives.
• Utilized Bayes' Theorem to update inlier probabilities based on hypothesis set verification, enabling parallel optimization.

Main Results:

• multiBaySAC demonstrated significantly higher computational efficiency (34% average improvement) compared to sequential RANSAC.
• The algorithm achieved high fitting accuracy, particularly in handling intersections between primitives.
• multiBaySAC effectively addressed over- and under-segmentation problems inherent in sequential RANSAC.

Conclusions:

• multiBaySAC offers a robust and efficient solution for fitting multiple primitives in 3D point clouds.
• The proposed parallel strategy and Bayesian probability updates enhance performance over traditional methods.
• Future research will explore multiBaySAC for applications like point cloud co-registration and image matching.