Surface Height and Uncertainty-Based Depth Estimation for Monocular 3D Object Detection

Monocular 3D object detection is a fundamental but challenging task in autonomous driving and robot navigation. Directly predicting the depth from a single image is essentially an ill-posed problem. Geometry projection is a powerful depth estimation method which infers an object’s depth from its physical height and projected height in the image plane. However, the height estimation errors will be amplified in the depth error. In this paper, we propose to estimate the physical heights and projected heights of object surface points instead of the height of the object itself so that a bunch of depth candidates are obtained. Besides, we estimate the uncertainties of the heights and replace the inferred instance depth of high uncertainty with the directly regressed depth. The final object depth is obtained by assembling these depth predictions according to the uncertainties. The experiments demonstrate the effectiveness of our depth estimation and our method achieves state-of-the-art results on the monocular 3D object detection task of the KITTI and Waymo open datasets.