Convex optimization for nonrigid stereo reconstruction.

Area of Science:

• Computer Vision
• 3-D Reconstruction
• Geometric Modeling

Background:

• Recovering 3-D nonrigid structure from stereo images is challenging.
• Existing methods often require smoothness constraints or initial estimations.
• Accurate reconstruction of nearly inextensible deformable surfaces is an open problem.

Purpose of the Study:

• To develop a robust and efficient method for 3-D nonrigid structure recovery from a single stereo image pair.
• To reconstruct surfaces with complex deformations, including sharp and smooth changes.
• To overcome limitations of existing methods by avoiding smoothness constraints and initial estimations.

Main Methods:

• Representing surfaces as 3-D triangulated meshes.
• Formulating reconstruction as an optimization problem with data and shape terms.
• Using second-order cone programming (SOCP) to handle keypoint correspondences (L-infinity norm).
• Approximating non-convex shape constraints (edge length preservation) as a sequence of convex SOCP feasibility problems.

Main Results:

• Successfully recovered 3-D nonrigid structures from stereo image pairs.
• Demonstrated accurate reconstruction of surfaces with smooth, sharp, and complex deformations.
• Achieved reliable and efficient solutions using an efficient SOCP solver.
• Validated robustness and accuracy on both synthetic and real-world data.

Conclusions:

• The proposed method effectively reconstructs 3-D nonrigid structures from stereo images.
• The approach is versatile, handling a wide range of surface deformations.
• Eliminating the need for smoothness constraints and initial estimations broadens applicability.
• The method offers a significant advancement in deformable surface reconstruction.