FCOS is a fully convolutional one-stage object detector that eliminates the need for anchor boxes and proposals, offering a simpler and more flexible approach to object detection. Unlike traditional detectors such as RetinaNet, SSD, and Faster R-CNN, which rely on predefined anchor boxes, FCOS predicts bounding boxes directly from each pixel in the feature map, similar to semantic segmentation. This approach avoids complex computations related to anchor boxes and reduces the number of hyperparameters, leading to a simpler and more efficient detector. FCOS achieves a 44.7% AP with a single model and single-scale test, surpassing previous one-stage detectors. The detector uses a "center-ness" branch to suppress low-quality bounding boxes, improving detection accuracy. FCOS also demonstrates strong performance as a Region Proposal Network (RPN) in two-stage detectors, outperforming anchor-based RPNs. The method is flexible, easy to implement, and can be extended to other vision tasks such as instance segmentation and keypoint detection. FCOS shows that anchor-free detectors can achieve state-of-the-art performance, challenging the traditional reliance on anchor boxes in object detection.FCOS is a fully convolutional one-stage object detector that eliminates the need for anchor boxes and proposals, offering a simpler and more flexible approach to object detection. Unlike traditional detectors such as RetinaNet, SSD, and Faster R-CNN, which rely on predefined anchor boxes, FCOS predicts bounding boxes directly from each pixel in the feature map, similar to semantic segmentation. This approach avoids complex computations related to anchor boxes and reduces the number of hyperparameters, leading to a simpler and more efficient detector. FCOS achieves a 44.7% AP with a single model and single-scale test, surpassing previous one-stage detectors. The detector uses a "center-ness" branch to suppress low-quality bounding boxes, improving detection accuracy. FCOS also demonstrates strong performance as a Region Proposal Network (RPN) in two-stage detectors, outperforming anchor-based RPNs. The method is flexible, easy to implement, and can be extended to other vision tasks such as instance segmentation and keypoint detection. FCOS shows that anchor-free detectors can achieve state-of-the-art performance, challenging the traditional reliance on anchor boxes in object detection.