SNOA-Pose: Symmetry-Aware Refinement for Sim-to-Real Category-Level 6D Object Pose Estimation

Abstract

Category-level 6D object pose estimation recovers the rotation, translation, and scale of unseen object instances within specific categories from RGB-D observations. Voting-based methods such as CPPF++ can be trained without real pose annotations. However, the gap between clean CAD-based training samples and noisy, incomplete real RGB-D observations degrades correspondence prediction and vote aggregation, leaving residual errors in the generated pose candidates. For rotationally symmetric objects, the symmetry-invariant alignment objective further leaves translation components orthogonal to the symmetry axis weakly constrained. Moreover, selecting between visual-and geometric-stream candidates solely by post-alignment geometric loss discards the confidence accumulated during voting. To solve these problems, we propose SNOA-Pose, a coarse-to-fine framework built on the CPPF++ probabilistic dual-stream voter. SNOA first uses a compact MiniPointNet to correct each candidate from partial observation–canonical correspondences and then initializes online alignment optimization. For rotationally symmetric categories, the refined translation is retained while the alignment rotation is adopted, preserving a stable translation anchor. An Epistemic-Geometric Dual Arbitrator (EGDA) subsequently gates candidates with post-alignment geometric consistency and resolves the remaining choice with symmetry-aware direction-vote evidence. On NOCS REAL275, SNOA-Pose improves the reproduced CPPF++ baseline on all five reported category-averaged metrics and reduces the post-processing time by 71.5% when using 30 alignment steps. An offline parallel-gripper proxy further shows that millimeter-level translation improvements produce larger gains under strict task tolerances.