Anchor-free detection is an approach to object detection that predicts bounding boxes directly from feature map locations, without relying on pre-defined anchor boxes. Traditional detectors like Faster R-CNN and early YOLO versions tile thousands of anchor boxes across the image at multiple scales and aspect ratios, then classify and refine each one. This works, but it introduces hyperparameters (anchor sizes, ratios, IoU thresholds) that need manual tuning per dataset.

Anchor-free methods skip this entirely. Keypoint-based detectors like CornerNet and CenterNet predict object corners or center points on a heatmap. Per-pixel methods like FCOS assign every feature map location inside a ground-truth box as a positive sample and regress distances to the four box edges. This simplifies the training pipeline and removes anchor-related hyperparameters.

The trend in 2024-2026 has moved further: architectures like DETR, RT-DETR, and YOLO26 use learned object queries or one-to-one label assignment, eliminating both anchors and NMS post-processing. These end-to-end designs reduce inference latency and produce more consistent results across different object sizes.