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

Author here. D-FINE-seg is a real-time vision framework built on D-FINE (a DETR variant). One architecture, one config flag (task: detect | segment | sem_seg), five sizes N/S/M/L/X - switches between object detection, instance segmentation, and semantic segmentation.

The most recent part I built: the semantic seg head that outputs a dense mask, you can think of it as a per-pixel classification. No queries, no NMS, just a small conv neck + 1x1 classifier on full-frame features. It also exports to ONNX / TensorRT / OpenVINO / CoreML / LiteRT with a per-backend parity check.

On Cityscapes (TensorRT FP16, RTX 5070 Ti) it comes out ahead of YOLO26 and RF-DETR on detection/instance-segmentation F1 and takes the top mIoU on semantic segmentation, at real-time latency with 2–3x fewer params. I tried hard to make the benchmark difficult to game: every framework runs its own shipped inference code, each gets its own confidence threshold tuned for F1, and a single validator scores all of them against the same GT. Full protocol and every known asymmetry is public here: https://github.com/ArgoHA/cityscapes-benchmark.

It's not a fork - the detection core follows the D-FINE paper, but the segmentation heads, training, export and inference are implemented from scratch. You can take the code and train/finetune on your dataset.

Blog write-up on the semantic segmentation work: https://blog.agnify.ai/d-fine-seg-semantic-segmentation/. Happy to take criticism on the eval.