CenterTrack applies a detection model to pairs of images and prior-frame detections to simultaneously localize objects and predict their associations across frames. It achieves 67.3% MOTA on MOT17 at 22 FPS and 89.4% MOTA on KITTI, setting state-of-the-art results on both benchmarks. The method also extends to monocular 3D tracking by regressing additional spatial attributes, achieving 28.3% AMOTA@0.2 on nuScenes at 28 FPS.