Inextensible Non-Rigid Structure-from-Motion by Second-Order Cone Programming.

Area of Science:

• Computer Vision and Machine Learning
• 3D Reconstruction and Motion Estimation

Background:

• Non-Rigid Structure-from-Motion (NRSfM) is challenging due to complex, nonlinear deformations.
• Existing methods often use affine cameras and factorization, limiting accuracy with intricate deformations.
• Template-less 3D reconstruction requires robust handling of deforming objects from image sequences.

Purpose of the Study:

• To develop a global and convex formulation for template-less 3D reconstruction of deforming objects.
• To adapt Second-Order Cone Programming (SOCP) for Non-Rigid Structure-from-Motion (NRSfM) using the Maximum-Depth Heuristic (MDH).
• To provide a robust and temporally continuous NRSfM solution that handles real-world complexities.

Main Methods:

• Formulation of a Second-Order Cone Programming (SOCP) problem for NRSfM based on the Maximum-Depth Heuristic (MDH).
• Maximizing point depths under geodesic distance constraints to enable convex optimization.
• Incorporation of robustness for outlier handling and temporal continuity for efficiency.

Main Results:

• The first convex formulation for NRSfM incorporating physical constraints.
• Accurate reconstruction of quasi-isometric objects from partial views, even with articulated and strong deformations.
• Superior or comparable accuracy to existing methods, with natural handling of missing correspondences and non-smooth objects.

Conclusions:

• The proposed SOCP-based MDH method offers a significant advancement in template-less 3D reconstruction for deforming objects.
• The formulation provides a robust, accurate, and simpler alternative to traditional NRSfM techniques.
• This approach opens new possibilities for analyzing and reconstructing dynamic 3D scenes.