FCOS: Fully Convolutional One-Stage Object Detection

FCOS is a fully convolutional one-stage object detector that reformulates detection as a per-pixel prediction problem, analogous to semantic segmentation. Whereas nearly all state-of-the-art detectors, including RetinaNet, SSD, YOLOv3, and Faster R-CNN, rely on pre-defined anchor boxes, FCOS is both anchor-box free and proposal free. Removing anchor boxes lets it avoid the complicated computation of box overlaps during training and, importantly, eliminates all the anchor-related hyperparameters that are often highly sensitive to final detection performance.

With non-maximum suppression as its only post-processing step, FCOS using a ResNeXt-64x4d-101 backbone achieves 44.7% AP under single-model, single-scale testing, surpassing previous one-stage detectors while remaining considerably simpler. The authors present this as the first demonstration that a much simpler and more flexible detection framework can achieve improved accuracy, and suggest FCOS can serve as a strong alternative baseline for other instance-level recognition tasks.