The Manhattan Frame Model-Manhattan World Inference in the Space of Surface Normals.

Area of Science:

• Computer Vision
• Robotics
• Geometric Modeling

Background:

• Man-made environments exhibit organized structures with orthogonal and parallel planes.
• Traditional Manhattan World (MW) assumptions are locally restrictive.
• Surface normal distributions reveal orthogonal clusters reflecting scene organization.

Purpose of the Study:

• Introduce the Manhattan-Frame (MF) model to capture MW properties in surface normals.
• Develop probabilistic MF models for single frames and mixtures (MMF).
• Enable robust inference and rotation estimation in man-made environments.

Main Methods:

• Proposed real-time MAP inference algorithms for single MF models.
• Developed MAP inference and adaptive MCMC sampling for Mixture of Manhattan Frames (MMF).
• Utilized Metropolis-Hastings split/merge moves for inferring mixture components.

Main Results:

• Demonstrated real-time performance and drift-free rotation estimation using single MF models.
• Showcased the versatility of the MMF model across various scales of man-made scenes.
• Successfully inferred scene structure and the number of mixture components.

Conclusions:

• The MF and MMF models provide a powerful framework for analyzing man-made environments.
• Proposed inference algorithms are effective for both single and complex scene structures.
• This work advances geometric reasoning in computer vision and robotics.