Cityscapes: D-FINE-seg vs YOLO26 vs RF-DETR

A like-for-like benchmark of three detection/segmentation frameworks on Cityscapes, across three tasks: detection, instance segmentation, and semantic segmentation.

• PROTOCOL.md — how every model was trained, exported, timed and scored, plus a full list of the asymmetries that remain.
• RESULTS.md / RESULTS.csv — the numbers.

The rules this benchmark plays by

1. Each framework runs its own shipped inference code. YOLO(engine).predict() for ultralytics, TRTInference + their PostProcess for RF-DETR, TRT_model for D-FINE. We never hand-write a rival's pre/postprocessing — we compare architecture + the framework it ships in, because that is what a user actually gets.
2. Accuracy and latency come from the same run, at each model's own val-optimal confidence threshold. Reporting F1 at one threshold and latency at another is not a benchmark.
3. Two latency columns: the framework's end-to-end call, and pure TensorRT engine time. They disagree — the rivals' CPU preprocessing dominates end-to-end — and reporting only one of them would be misleading.
4. Sizes are matched on latency, not parameters. Each rival is the smallest variant with at least D-FINE's engine budget, so it is never under-resourced (it also carries 2–3× the params).
5. One scorer for everyone: D-FINE's Validator, original-resolution ground truth. It reproduces D-FINE's own bench.py output exactly — which is what makes the rival rows trustworthy.
6. Every asymmetry we know about is written down in PROTOCOL.md §9, including the ones that don't flatter us.

Quick start

python prep_datasets.py                          # dataset layouts (symlinks, original resolution)
./run_yolo_training.sh                           # train YOLO26
./run_rfdetr_training.sh                         # train RF-DETR
python bench_all.py --models dfine yolo rfdetr   # bench: accuracy + latency at the same operating point
python make_results.py                           # regenerate RESULTS.md / RESULTS.csv

Requires three isolated venvs (one per framework), each pinned to TensorRT 10.13.3.9 so all engines are mutually comparable. Paths are hard-coded at the top of each script — edit them to run elsewhere.

Hardware for the published numbers: single RTX 5070 Ti (16 GB), TensorRT 10.13.3.9, FP16, batch 1, idle GPU.