Marginalized Bundle Adjustment:
Multi-View Camera Pose from Monocular Depth Estimates

We introduce Marginalized Bundle Adjustment (MBA), a bundle-adjustment objective tailored to monocular depth maps. MBA registers monocular depth maps into world coordinate. It accounts mono-depthmaps' high variance with its density. Excitingly, we show that monocular depth maps are already accurate enough to achieve state-of-the-art results on both small- and large-scale Structure-from-Motion and camera relocalization tasks.

MBA formulates the mono-depth's projective residual as an error distribution R={r}. Its forward function is indexing the Cumulative-Distribution-Function (CDF) of residual distribution R. Its backward function is indexing the Probability-Distribution-Function (PDF) of residual distribution R. The Bundle-Adjustment operates to maximize the Area-Under-the-Curve (AUC) of CDF function.