retinanet

RetinaNet (Lin et al., 2017) is the one-stage detector that proved a dense, single-pass model could be as accurate as the best two-stage detectors, not just faster. Before it, the rule of thumb was 'one-stage for speed, two-stage for accuracy'. RetinaNet broke that trade-off by combining a clean multi-scale architecture with focal loss, and in doing so it became the standard reference one-stage detector for years.

Architecturally it is deliberately simple: a backbone (e.g. ResNet) plus a feature pyramid network produce multi-scale features P3–P7, and two small fully-convolutional subnets are applied identically to every pyramid level — one for classification (predicting K class probabilities per anchor) and one for box regression (4 offsets per anchor). Anchors of multiple scales and aspect ratios tile each level densely. There is no proposal stage and no second refinement; everything is predicted in one sweep.

The crucial ingredient is the loss. Such a dense design evaluates ~100k anchors per image, overwhelmingly background — the exact imbalance that had always held one-stage detectors back. RetinaNet trains the classification subnet with focal loss, which down-weights the easy negatives so the rare hard positives drive learning. This single change, more than the architecture, is why RetinaNet works; the paper's thesis is essentially 'the obstacle to one-stage accuracy was class imbalance, and focal loss removes it'.

RetinaNet on COCO surpassed contemporary two-stage detectors (including Faster R-CNN with FPN) while running faster, and it crystallized the now-standard one-stage recipe — backbone + FPN + shared dense heads + a strong imbalance-aware loss — that later anchor-free detectors (FCOS) and many YOLO versions adopted. It remains a common baseline because it is conceptually clean and easy to reason about.