Ultralytics Benchmark#

from ultralytics import YOLO

datasets = [
    "ul://ultralytics/datasets/airbus-oil-storage-detection",
    "ul://ultralytics/datasets/annopage-doc-elements",
    "ul://ultralytics/datasets/basketball-hoop-detection",
    "ul://ultralytics/datasets/ceymo",
    "ul://ultralytics/datasets/cheng-brain-tumor-mri",
    "ul://ultralytics/datasets/ddr-lesion",
    "ul://ultralytics/datasets/deeppcb",
    "ul://ultralytics/datasets/dentex",
    "ul://ultralytics/datasets/dspcbsd-pcb-surface-defect",
    "ul://ultralytics/datasets/egohands-egocentric",
    "ul://ultralytics/datasets/ena24",
    "ul://ultralytics/datasets/fisheye8k-traffic",
    "ul://ultralytics/datasets/floorplancad-2",
    "ul://ultralytics/datasets/grazpedwri-dx",
    "ul://ultralytics/datasets/groundcua-ui-yolo",
    "ul://ultralytics/datasets/global-wheat-head-detection-2021",
    "ul://ultralytics/datasets/hit-uav",
    "ul://ultralytics/datasets/iedxray",
    "ul://ultralytics/datasets/kvasir-seg-polyp",
    "ul://ultralytics/datasets/llvip",
    "ul://ultralytics/datasets/moda-multispectral-obb",
    "ul://ultralytics/datasets/plantdoc",
    "ul://ultralytics/datasets/racketvision-tennis-ball-yolo",
    "ul://ultralytics/datasets/rdd2022",
    "ul://ultralytics/datasets/rpc-retail-checkout",
    "ul://ultralytics/datasets/ruod-underwater",
    "ul://ultralytics/datasets/safe-human-17",
    "ul://ultralytics/datasets/seadronessee-2",
    "ul://ultralytics/datasets/sku110k-2",
    "ul://ultralytics/datasets/soccerdb",
    "ul://ultralytics/datasets/ssdd-sar-ship",
    "ul://ultralytics/datasets/taco-detect",
    "ul://ultralytics/datasets/tbx11k-tb-only",
    "ul://ultralytics/datasets/tn5000-thyroid-us",
    "ul://ultralytics/datasets/txl-pbc-blood-cells",
    "ul://ultralytics/datasets/uec-school-lunch-food-2",
    "ul://ultralytics/datasets/visdrone",
]

results = YOLO("yolo27n.pt").train(
    data=datasets,
    epochs=100,
    patience=100,
    imgsz=640,
    batch=16,
    fraction=[1000, 1.0, 0.0],
    seed=0,
    workers=4,
    cls_remap=True,
    channels_last=True,
    save_json=True,
    plots=False,
)
# Der Haupt-Branch verwendet Größenkategorien mit nativer Fläche. Dieser Benchmark skaliert GT- und Vorhersageflächen separat
# mit (640 / max(image_width, image_height)) ** 2, bevor AP nach Größe berechnet wird.

@misc{jocher2026ultralyticsyolo26unifiedrealtime,
  title = {Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models},
  author = {Glenn Jocher and Jing Qiu and Mengyu Liu and Shuai Lyu and Fatih Cagatay Akyon and Muhammet Esat Kalfaoglu},
  year = {2026},
  eprint = {2606.03748},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  doi = {10.48550/arXiv.2606.03748},
  url = {https://arxiv.org/abs/2606.03748},
}