What is anchor-free object detection and how does it differ from anchor-based methods?

Anchor-free detectors predict object centers or keypoints instead of predefined anchor boxes.

Anchor-based methods like Faster R-CNN rely on predefined bounding boxes, which require tuning scales and aspect ratios. Anchor-free methods like CenterNet and FCOS directly predict object centers, bounding box distances, or keypoints, simplifying design and improving generalization across datasets.