Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection

Monocular 3D object detection aims to localize 3D bounding boxes in an input\nsingle 2D image. It is a highly challenging problem and remains open,\nespecially when no extra information (e.g., depth, lidar and/or multi-frames)\ncan be leveraged in training and/or inference. This paper proposes a simple yet\neffective formulation for monocular 3D object detection without exploiting any\nextra information. It presents the MonoCon method which learns Monocular\nContexts, as auxiliary tasks in training, to help monocular 3D object\ndetection. The key idea is that with the annotated 3D bounding boxes of objects\nin an image, there is a rich set of well-posed projected 2D supervision signals\navailable in training, such as the projected corner keypoints and their\nassociated offset vectors with respect to the center of 2D bounding box, which\nshould be exploited as auxiliary tasks in training. The proposed MonoCon is\nmotivated by the Cramer-Wold theorem in measure theory at a high level. In\nimplementation, it utilizes a very simple end-to-end design to justify the\neffectiveness of learning auxiliary monocular contexts, which consists of three\ncomponents: a Deep Neural Network (DNN) based feature backbone, a number of\nregression head branches for learning the essential parameters used in the 3D\nbounding box prediction, and a number of regression head branches for learning\nauxiliary contexts. After training, the auxiliary context regression branches\nare discarded for better inference efficiency. In experiments, the proposed\nMonoCon is tested in the KITTI benchmark (car, pedestrain and cyclist). It\noutperforms all prior arts in the leaderboard on car category and obtains\ncomparable performance on pedestrian and cyclist in terms of accuracy. Thanks\nto the simple design, the proposed MonoCon method obtains the fastest inference\nspeed with 38.7 fps in comparisons\n