Reproduce “CSS-BA: Gate-Guided Column Space Search for Bundle Adjustment” on OpenResearch

About this paper

Ayano Kaneda, Takafumi Taketomi, Shugo Yamaguchi, Shigeo Morishima

Gate-Guided CSS-BA introduces a modification for Bundle Adjustment (BA) that stabilizes camera pose and calibration estimation, particularly in challenging weak-geometry conditions such as low-parallax or near-rotational camera movements. This method improves relative pose accuracy and reduces focal-length errors by restricting the Levenberg-Marquardt update to a geometrically informed, low-dimensional search subspace.

Key insights

• Directly stabilizing the Levenberg-Marquardt update direction by restricting it to a geometrically informed low-dimensional subspace can significantly improve BA robustness in ill-conditioned settings without altering the core problem formulation.
• Pre-filtering cameras using a 'geometric gate' based on standard SfM/SLAM diagnostics effectively identifies and prioritizes well-conditioned camera blocks, guiding the subspace search towards stable regions.
• Constructing a search basis using Ritz vectors derived from a localized Schur operator on selected camera blocks allows the solver to focus on dominant curvature directions, avoiding poorly supported updates along near-null directions.

Summary by alphaXiv.