Triangulation-Free Bundle Adjustment with Graduated Non-Convexity for Camera Pose Refinement from Coarse Priors

An AR capture provides a metric camera pose for every frame, accurate to a fraction of a degree, which is not accurate enough for novel-view synthesis. The common approach triangulates structure from the prior and bundle-adjusts, which commits the prior’s error into the structure the optimiser then trusts, and can drop frames that fail to register. This work refines the prior directly: each keypoint carries one scalar depth along its own back-projected ray, each match contributes two symmetric cross-projection residuals, and structure is re-expressed at every iterate rather than fixed in advance. Because no structure is fixed, the objective can be annealed under graduated non-convexity; because there is no reconstruction step, every input pose is refined and no frame is dropped.