Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection

Monocular 3D object detection aims to localize bounding boxes in an input single 2D image. It is a highly challenging problem and remains open, especially when no extra information (e.g., depth, lidar and/or multi-frames) can be leveraged training inference. This paper proposes simple yet effective formulation for monocular without exploiting any information. presents the MonoCon method which learns Contexts, as auxiliary tasks training, help detection. The key idea that with annotated of objects image, there rich set well-posed projected supervision signals available such corner keypoints their associated offset vectors respect center box, should exploited training. proposed motivated by Cramer–Wold theorem measure theory at high level. In implementation, it utilizes very end-to-end design justify effectiveness learning contexts, consists three components: Deep Neural Network (DNN) based feature backbone, number regression head branches essential parameters used box prediction, contexts. After context are discarded better inference efficiency. experiments, tested KITTI benchmark (car, pedestrian cyclist). outperforms all prior arts leaderboard on car category obtains comparable performance cyclist terms accuracy. Thanks design, fastest speed 38.7 fps comparisons. Our code released https://git.io/MonoCon.