Projective multiview structure and motion from element-wise factorization.

Area of Science:

• Computer Vision
• Robotics
• Computational Geometry

Background:

• Traditional iterative methods like Sturm-Triggs for projective structure-from-motion (SfM) factorization face challenges with initialization, convergence, and local minima.
• These iterative algorithms struggle with scalability and handling real-world data imperfections such as missing data and outliers.

Purpose of the Study:

• To reformulate the projective SfM problem using element-wise (Hadamard) factorization instead of conventional matrix factorization.
• To develop a novel convex optimization approach for simultaneously solving projective depths, scene structure, and camera motions.
• To address scalability issues and enhance robustness to missing data and outliers in SfM.

Main Methods:

• Formulation of the projective SfM problem as an element-wise (Hadamard) factorization.
• Application of convex optimization techniques to solve for depths, structure, and motions simultaneously.
• Adoption of a continuation-based algorithm to manage scalability.

Main Results:

• The proposed method achieves a globally optimal solution for the projective SfM problem, up to a relaxation gap.
• Demonstrated ability to handle missing data and outliers in a unified and natural manner.
• Experimental results on synthetic and real images show performance comparable to state-of-the-art methods.

Conclusions:

• The novel element-wise factorization approach provides a robust and globally optimal solution for projective structure-from-motion.
• Convex optimization offers a powerful alternative to iterative methods, improving convergence and handling of complex data.
• The method's effectiveness in real-world scenarios highlights its potential for advanced computer vision and robotics applications.