Boosting Monocular 3D Object Detection With Object-Centric Auxiliary Depth Supervision

Recent advances in monocular 3D detection leverage a depth estimation network explicitly as an intermediate stage of the network. Depth map approaches yield more accurate to objects than other methods thanks trained on large-scale dataset. However, can be limited by accuracy map, and sequentially using two separated networks for significantly increases computation cost inference time. In this work, we propose method boost RGB image-based detector jointly training with prediction loss analogous task. way, our supervised supervision from raw LiDAR points, which does not require any human annotation cost, estimate without predicting map. Our novel object-centric focuses around foreground objects, is important object detection, pixel-wise manner. regression model further predict uncertainty represent confidence objects. To effectively train points enable end-to-end training, revisit target design architecture. Extensive experiments KITTI nuScenes benchmarks show that outperform while maintaining real-time speed.