Multiview Point Cloud Registration via Optimization in an Autoencoder Latent Space.

Area of Science:

• 3D Computer Vision
• Computational Geometry
• Robotics

Background:

• Point cloud registration is crucial for 3D reconstruction and analysis.
• Existing multiview registration methods struggle with scalability and robustness to degradations.
• Generative models are limited by Gaussian Mixture Models and Expectation-Maximization, hindering large transformations.

Purpose of the Study:

• Introduce POLAR (Point cloud Latent Registration) for efficient and robust multiview point cloud registration.
• Address limitations of existing methods in handling numerous views and significant initial errors.
• Develop a method resilient to high levels of data degradation.

Main Methods:

• Transposing the registration problem into the latent space of a pre-trained autoencoder.
• Designing a specialized loss function to account for data degradations.
• Implementing an efficient multistart optimization strategy for improved convergence.

Main Results:

• POLAR demonstrates superior performance compared to state-of-the-art methods.
• The method achieves significant improvements on both synthetic and real-world datasets.
• POLAR exhibits robustness to a high number of views and large initial transformation angles.

Conclusions:

• POLAR offers a scalable and robust solution for multiview point cloud registration.
• The latent space approach effectively handles complex registration scenarios.
• The method advances the state-of-the-art in 3D computer vision applications.