Joint Alignment of Multiple Point Sets with Batch and Incremental Expectation-Maximization.

Area of Science:

• Computer Vision
• Computational Geometry
• Machine Learning

Background:

• Traditional pairwise registration methods treat point sets unevenly, leading to noise and outlier sensitivity.
• Existing techniques can be suboptimal when dealing with multiple, potentially noisy, 3D point sets.

Purpose of the Study:

• To develop a unified and robust method for registering multiple point sets.
• To overcome the limitations of pairwise registration by treating all sets symmetrically.

Main Methods:

• Formulating point set registration as a Gaussian Mixture Model (GMM) clustering problem.
• Deriving batch and incremental Expectation-Maximization (EM) algorithms for robust parameter estimation.
• Utilizing GMM means for aligned point sets and variances for alignment contribution insights.

Main Results:

• The proposed EM algorithms robustly estimate GMM parameters, rotations, and translations.
• Demonstrated superior performance compared to state-of-the-art methods on simulated and real-world range sensor data.
• Showcased potential for accurate surface reconstruction from depth data.

Conclusions:

• The GMM-based clustering approach offers a symmetrical and robust solution for multi-set registration.
• The method effectively handles noise and outliers, improving alignment accuracy.
• This technique shows significant promise for applications like 3D surface reconstruction.