D-FINE-seg Unifies Object Detection and Segmentation in One Real-Time Model

D-FINE-seg – detection, instance and semantic segmentation in one model

I built D-FINE-seg to handle object detection, instance segmentation, and semantic segmentation within a single codebase. This framework beats YOLO26 and RF-DETR on Cityscapes benchmarks with significantly fewer parameters while maintaining real-time latency. It supports modern training features like the Muon optimizer and exports to multiple backends including TensorRT and OpenVINO, all configured with just one flag.