yolo26m-doc-layout-onnx

yolo26m-doc-layout-onnx

An ONNX export of Armaggheddon/yolo26-document-layout (yolo26m_doc_layout.pt), a DocLayNet-trained document-layout detector. All credit for the weights goes to that repository; this one only changes the container format.

Why the export exists. The weights are MIT, but the runtime that loads a .pt is ultralytics, which is AGPL-3.0 and cannot ship inside a proprietary image. The ONNX graph runs on onnxruntime (MIT), so exporting once separates the license of the model from the license of the loader. Nothing downstream needs ultralytics or torch.

Files

sha256  b143fbb74dbd24ae5b7ebf7b94e78983e02660e4175242b2d2a0f83575848373

Use 1280. These weights were trained at 1280 and degrade at the ultralytics default of 640. The input shape is pinned in the graph, so a mismatched letterbox is a runtime error at best and silently rescaled boxes at worst.

Classes

The export carries its names metadata, so class ids map to labels without being told the order — 11 DocLayNet classes:

Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, Title

Usage

curl -fsSL -o yolo26m_doc_layout.onnx \
  https://huggingface.co/superchaintech/yolo26m-doc-layout-onnx/resolve/main/yolo26m_doc_layout.onnx
echo "b143fbb74dbd24ae5b7ebf7b94e78983e02660e4175242b2d2a0f83575848373  yolo26m_doc_layout.onnx" | sha256sum -c

import numpy as np, onnxruntime
from PIL import Image

session = onnxruntime.InferenceSession(
    "yolo26m_doc_layout.onnx", providers=["CPUExecutionProvider"],
)
# letterbox your page to 1280x1280, CHW, float32 /255, then:
# outputs = session.run(None, {session.get_inputs()[0].name: batch})

With sc_toolkit, which wraps letterboxing, NMS and the label mapping:

import os
os.environ["SC_LAYOUT_MODEL_DIR"] = "/opt/layout"
from sc_toolkit.vlm.hybrid.detectors import load_onnx_detector

detector = load_onnx_detector("yolo26m_doc_layout.onnx")
regions = detector.detect(page_image, 0.25)   # 0.25 is the YOLO calibration

min_conf matters: 0.25 is the calibration for this backend. 0.5 is the PP-DocLayoutV3 number and drops span coverage to ~79%.

Measured behaviour

Six real pages (a KPMG accounting paper and a Hong Kong annual report) at 150 dpi, min_conf=0.25: 78 regions found, no empty page, top confidences 0.84–0.99.

Reproducing the export

pip install ultralytics huggingface_hub onnx
python -c "from huggingface_hub import hf_hub_download; from ultralytics import YOLO; \
  w = hf_hub_download('Armaggheddon/yolo26-document-layout', 'yolo26m_doc_layout.pt'); \
  print(YOLO(w).export(format='onnx', imgsz=1280, opset=17, simplify=True))"