The paper introduces FCOS (Fully Convolutional One-Stage Object Detector), a novel anchor-free and proposal-free object detection method. Unlike traditional detectors that rely on pre-defined anchor boxes, FCOS predicts bounding boxes directly at each pixel, similar to semantic segmentation tasks. This approach eliminates the need for complex anchor box calculations and hyper-parameter tuning, making the detection process simpler and more efficient. The key innovation is the "center-ness" branch, which helps suppress low-quality predictions by penalizing boxes far from the center of objects. Experiments on the COCO dataset show that FCOS achieves 44.7% AP with single-model and single-scale testing, outperforming state-of-the-art one-stage detectors like RetinaNet and CornerNet. The paper also demonstrates that FCOS can be used as a Region Proposal Network (RPN) in the two-stage detector Faster R-CNN, further improving performance. Overall, FCOS provides a robust and flexible alternative to anchor-based detectors, with simpler architecture and better detection accuracy.The paper introduces FCOS (Fully Convolutional One-Stage Object Detector), a novel anchor-free and proposal-free object detection method. Unlike traditional detectors that rely on pre-defined anchor boxes, FCOS predicts bounding boxes directly at each pixel, similar to semantic segmentation tasks. This approach eliminates the need for complex anchor box calculations and hyper-parameter tuning, making the detection process simpler and more efficient. The key innovation is the "center-ness" branch, which helps suppress low-quality predictions by penalizing boxes far from the center of objects. Experiments on the COCO dataset show that FCOS achieves 44.7% AP with single-model and single-scale testing, outperforming state-of-the-art one-stage detectors like RetinaNet and CornerNet. The paper also demonstrates that FCOS can be used as a Region Proposal Network (RPN) in the two-stage detector Faster R-CNN, further improving performance. Overall, FCOS provides a robust and flexible alternative to anchor-based detectors, with simpler architecture and better detection accuracy.