anchor-free detection

Anchor-free detection predicts objects directly from image locations — points, centers, or corners — without first laying down a grid of predefined anchor boxes. The motivation is to delete the whole headache that anchors bring: their scales and aspect ratios are hyperparameters you must tune to your data, they create a huge imbalance of mostly-background boxes, and matching anchors to objects by IoU is a fiddly, dataset-specific recipe. If you can detect objects without ever defining an anchor, all of that disappears.

There are two main flavors. Keypoint-based methods represent an object by special points and group them: CornerNet predicts top-left and bottom-right corner heatmaps and pairs corners belonging to the same object via learned embeddings; CenterNet (the 'Objects as Points' variant) goes simpler, predicting a single center-point heatmap per class plus regressed width/height and a small offset at each peak — detection becomes keypoint estimation with no NMS in the box sense. ExtremeNet uses extreme points (topmost, leftmost, etc.).

Center/dense methods keep the per-pixel dense-prediction style of RetinaNet but drop anchors. FCOS (Fully Convolutional One-Stage) treats every feature-map location that falls inside a ground-truth box as a positive sample and, at that location, directly regresses the four distances to the box's left/right/top/bottom edges, plus a class score and a 'center-ness' score that down-weights predictions far from the object's center (a learned substitute for NMS-friendly confidence). Multiple FPN levels resolve which scale of object each location is responsible for.

Anchor-free detectors match or beat anchor-based ones in accuracy while having fewer hyperparameters and often simpler code, which is why the modern YOLO line (v8 onward) and many production detectors went anchor-free. Note that most still rely on NMS to remove duplicate detections — 'anchor-free' removes the anchor prior, not necessarily the post-processing; fully removing NMS as well is the separate contribution of set-prediction transformers like DETR.