Objects as Points

This paper reframes object detection by modeling each object as a single point, the center of its bounding box, rather than enumerating a nearly exhaustive list of candidate boxes and classifying each, which the authors argue is wasteful, inefficient, and dependent on additional post-processing. Their detector, CenterNet, uses keypoint estimation to find center points and then regresses to all other object properties, including size, 3D location, orientation, and even pose. The approach is end-to-end differentiable and simpler than bounding-box-based detectors.

CenterNet achieves the best speed-accuracy trade-off on the MS COCO dataset, reaching 28.1% AP at 142 FPS, 37.4% AP at 52 FPS, and 45.1% AP with multi-scale testing at 1.4 FPS, while being faster and more accurate than comparable box-based methods. The same center-point approach extends to estimating 3D bounding boxes on the KITTI benchmark and human pose on the COCO keypoint dataset, performing competitively with sophisticated multi-stage methods while running in real time, which mattered as a clean and general point-based alternative to anchor-based detection.