Monocular 3D Object Detection Based on Occlusion Optimization

Obtaining accurate 3D bounding boxes from monocular images is a challenging task. In autonomous driving scenarios, mutual occlusion between objects is a common situation, and that is one of the main factors affecting the accuracy of 3D object detection. Aiming at solving the occlusion problem, this paper proposes a monocular 3D object detection method that integrates occlusion differential processing. We propose an occlusion processing branch to process the filtered occlusion feature maps from global and optimize the 3D prediction results with predicted occlusion compensation. We validate our method on the KITTI 3D object detection benchmark, and experiments show that our prediction accuracy for occluded objects is significantly improved.