This repository implements CenterMask, an anchor-free one-stage instance segmentation approach built on FCOS object detection. It adds a novel SAG-Mask branch that predicts segmentation masks using spatial attention maps to focus on informative pixels. The project includes both the full CenterMask model and a lightweight CenterMask-Lite variant designed for real-time inference, achieving state-of-the-art results faster than previous methods like Mask R-CNN and YOLACT.