FCOS

Why we included this project

FCOS showed that a one-stage detector can skip anchor boxes entirely and still beat classic region-based methods like Faster R-CNN. The fully convolutional head predicts bounding boxes directly from the feature map, so you never have to tune anchor hyperparameters. If you work in PyTorch and use FPN backbones, the codebase is straightforward and the repo includes the published recipe for reproducing strong COCO results. It also works well as a starting point for prototyping anchor-free detection or comparing against later detectors built on this design. For research-oriented engineers, this is a solid, well-documented baseline to fork and extend rather than a turnkey product.