This document summarizes a paper on CenterTrack, a new object tracking method. CenterTrack localizes objects as keypoints and associates them across frames with minimal input. It applies a detection model to pairs of images and prior detections to localize objects and predict associations. CenterTrack represents an improvement over prior methods by using keypoints instead of bounding boxes and requiring only previous detections as input for tracking. An ablation study showed that offset prediction and multi-scale feature maps were important for performance.