MonoPRIO: Adaptive Prior Conditioning for Unified Monocular 3D Object Detection
MonoPRIO addresses unstable metric size estimation in monocular 3D detection by introducing adaptive prior conditioning in the size pathway. It builds class-aware prototypes offline, routes decoder queries to these priors using soft mixtures, and applies uncertainty-aware log-space conditioning. This method achieves superior unified performance on KITTI, especially in ambiguous scenarios like occlusion.