Multi-task Cooperative Monocular 3D Object Detection Using Surface Center Keypoints

Detecting 3D objects from a single monocular image without spatial information is a challenging task. In this paper, we propose a 3D object detection model which directly regresses the center point, depth, size and orientation of the 3D bounding box using a single monocular RGB image. In particular, we obtain multiple surface center points of each 3D box in two ways, one of which is directly predicting them, and the other is indirectly obtaining them through geometry computation. The error between the results in two ways acts as a new unified loss item to improve the convergence of the prediction on surface center points. The proposed approach has been verified on the public nuScenes dataset and experiments demonstrate that the additional keypoints estimation and the unified loss improve some aspects of 3D object detection.