The paper presents CenterTrack, a novel framework for object detection and tracking that simplifies the traditional tracking pipeline by treating objects as points. CenterTrack combines the strengths of object detection and tracking, leveraging deep learning to perform both tasks simultaneously. The key idea is to use a detection model to localize objects in pairs of consecutive frames, incorporating a heatmap of prior tracklets as input. This approach allows the model to predict an offset vector from the current object center to its previous frame position, enabling efficient and accurate object association. CenterTrack is designed to be simple, online, and real-time, achieving high performance on multiple benchmarks, including MOT17 and KITTI, with state-of-the-art results. The method is also extended to monocular 3D tracking, outperforming existing methods on the nuScenes dataset. The paper discusses the design choices, training procedures, and ablation studies, highlighting the effectiveness of the proposed approach in handling complex tracking tasks.The paper presents CenterTrack, a novel framework for object detection and tracking that simplifies the traditional tracking pipeline by treating objects as points. CenterTrack combines the strengths of object detection and tracking, leveraging deep learning to perform both tasks simultaneously. The key idea is to use a detection model to localize objects in pairs of consecutive frames, incorporating a heatmap of prior tracklets as input. This approach allows the model to predict an offset vector from the current object center to its previous frame position, enabling efficient and accurate object association. CenterTrack is designed to be simple, online, and real-time, achieving high performance on multiple benchmarks, including MOT17 and KITTI, with state-of-the-art results. The method is also extended to monocular 3D tracking, outperforming existing methods on the nuScenes dataset. The paper discusses the design choices, training procedures, and ablation studies, highlighting the effectiveness of the proposed approach in handling complex tracking tasks.