Objects as Points
lastmod 2026-11-10created 2019-11-073 edits
Objects as Points is the CenterNet paper, which detects objects by representing each one as a single center point instead of a bounding box. It uses keypoint estimation to find each object's center on a heatmap and, at that point, jointly predicts properties such as size and offset through regression. Because it needs neither anchor boxes nor NMS post-processing, its structure is simple, and the same framework extends to 3D detection and pose estimation.
Documentation
- Objects as Points
- https://arxiv.org/abs/1904.07850v2