RANSAC Optimization for Category-level 6D Object Pose Estimation

The performance of category-level 6D object pose estimation is insufficient although the RANSAC algorithm has been adopted to remove outliers from the 3D-3D correspondence. In this paper, we optimize the Preemptive RANSAC algorithm to progressively find an inlier subset that has the smallest reprojection error. Experiments demonstrate that the optimized algorithm gets a more precise 6D pose in which hypotheses can be measured against each other and be revised to estimate a more precise pose. For the 10° &10cm metric on the REAL275 dataset, it achieves 43.8% and exceeds 64.0% of the NOCS method. On the Context-Aware MixEd ReAlity(CAMERA25) dataset it achieves 63.2% which exceeds 1.6% of NOCS.