Multi-Scale Grid Attention and Probabilistic Refinement for Accurate RoI-Based Monocular 3D Object Detection

Monocular 3D object detection remains a challenging task due to the inherent limitations of single-view depth perception. While conventional methods treat all parts of the Region of Interest (RoI) uniformly, we argue that different RoI regions hold varying importance for accurate detection. This study introduces the Multi-Scale Grid Attention (MSGA) mechanism to investigate the significance of RoI regions at multiple scales. Furthermore, we propose a novel probabilistic post-processing method to enhance detection robustness by effectively utilizing the probabilistic properties of depth estimation during inference. Our approach achieves state-of-the-art performance on the KITTI and Waymo datasets, demonstrating significant improvements in detection accuracy and robustness.