This repository implements CenterNet, a deep learning approach that models objects as single points (their bounding box centers) and uses convolutional neural networks to detect these center points. The model regresses object properties including size, 3D location, and pose from center point predictions. It achieves real-time performance on MS COCO (142 FPS at 28.1% AP) and competitive results on KITTI 3D detection and COCO keypoint estimation benchmarks.