Skip the anchor boxes—detect objects by their center points

What it does

CenterNet treats object detection as keypoint estimation: it locates the center point of each object and regresses directly to bounding box dimensions, 3D location, orientation, and even human pose. The same core network handles 2D detection on COCO, 3D bounding boxes on KITTI, and multi-person pose estimation without architectural contortions. Everything runs in a single feedforward pass, and notably, it drops the usual Non-Maximum Suppression post-processing step entirely.

The interesting bit

The clever twist is modeling objects as points rather than boxes. Instead of maintaining anchor boxes of every shape and size, the network predicts a heatmap of object centers and simply reads off width, height, and depth from the features at those peak locations. It is end-to-end differentiable and, according to the authors, simpler and faster than corresponding bounding-box-based counterparts.

Key highlights

• Achieves 37.4 COCO AP at 52 FPS with a DLA-34 backbone, or 28.1 AP at 142 FPS with ResNet-18.
• Reaches 45.1 AP on COCO test-dev with multi-scale testing using an Hourglass-104 backbone.
• Handles 3D detection and human pose estimation with minor modifications to the same framework.
• No NMS required; the pipeline is a single network forward pass.
• Includes webcam demos and a Python API for dropping the detector into other projects.

Verdict

Worth a look if you want a clean, unified intuition for detection, pose, and 3D estimation without anchor-box boilerplate. Skip it if you need a production-ready, maintained framework out of the box—this is a research artifact with a rich family of successor projects.