MonoPRIO: Adaptive Prior Conditioning for Unified Monocular 3D Object Detection

MonoPRIO is a monocular 3D object detector that uses offline size prototypes, soft prior routing, and training regularization to improve metric size estimation from a single image.

It addresses a key weakness in monocular 3D detection: predicting reliable object size and depth under occlusion, ambiguity, and class variation.

On KITTI, the method reports strong results in both unified multi-class and car-only settings, outperforming prior approaches such as MonoCLUE.

The biggest gains appear in difficult scenes, especially for occluded or hard-to-disambiguate objects like cars, pedestrians, and cyclists.