Expert-Consistency-Guided Proposal Refinement for 3D Object Detection from Sparse Radar Point Clouds

Sparse radar point clouds provide limited evidence for three-dimensional (3D) object detection, making weak proposals especially sensitive to ranking and suppression. Standard post-processing depends on the primary detector's scores and geometry and can therefore remove useful candidates. We introduce an expert-consistency-guided refinement framework that acts on detector outputs without changing the backbone. Candidate boxes are retained and reweighted according to agreement with an independently trained radar expert; supported boxes are then adjusted by local geometric voting. A separate high-performance route instead trains a localization-quality-aware confidence head with relaxed positive assignment and does not use the expert-consistency gate. The framework is evaluated on Astyx, MAN TruckScenes, V2X-Radar-V, and K-Radar with three training seeds. At a 3D intersection-over-union threshold of 0.50, the strict route improves average precision (AP) in all 12 matched dataset-seed comparisons, with a mean gain of 0.9726 AP. The high-performance route, whose gain comes from this quality-aware ranking branch rather than from expert-consistency gating, raises macro AP from 35.3552 to 39.1214, a gain of 3.7662 AP. These improvements are observations under the evaluated protocol: the strict voting parameters were selected using the same dataset-seed cells, and leave-one-dataset-out threshold selection yields 11/12 positive held-out differences. The strict route also shows metric-dependent confidence-calibration trade-offs and requires an additional expert detector and voting computation. The results therefore support protocol-specific gains, not a guarantee of improved reliability on unseen data.