ICC-Pose: Implicit Category Consensus for One-Stage RGB Category-Level 9D Pose Estimation

Abstract

Monocular RGB category-level 9D pose estimation recovers the rotation, translation, and size of unseen instances without test-time depth, masks, instance-specific CAD models, or instance-level shape priors. One-stage query-based pose heads provide a compact detector-to-pose interface, yet each query must resolve ambiguity in shape, scale, and viewpoint. We introduce ICC-Pose, which embeds category consensus as a latent residual within each pose query. Provisional class evidence retrieves learnable prototype tokens; query–prototype attention aggregates them, and a gate injects the residual before pose regression. Auxiliary geometry supervision makes the residual pose-aware during training. Its predictions are discarded at test time, preserving RGB-only one-stage inference. Under joint CAMERA25 and REAL275 training, ICC-Pose improves all five YOPO metrics, with the largest gain of 6.35 points at 10$^\circ$/10 cm. When ICC-Pose and UniDP are evaluated with the released YOPO evaluator, ICC-Pose outperforms UniDP on four of five metrics. On HouseCat6D, it improves all six reported metrics, led by an 8.2-point gain at IoU50. A fixed-start, matched-budget comparison shows higher three-seed mean performance from geometry-supervised ICC refinement. A fixed-checkpoint intervention measures the deployed fusion path. Prototype diagnostics reveal a diverse bank whose attention concentration tracks reliability.