Guide

Appearance

YOLO26detectONNX Setup ​

_YOLO26detectONNX runs Ultralytics YOLO26 detection ONNX models with ONNX Runtime for inference and OpenCV for image preprocessing and NMS fallback. It supports the default YOLO26 end-to-end output (N, 300, 6) and the traditional one-to-many export output (N, nc + 4, predictions).

Dependencies ​

Install the normal OpenKAI OpenCV dependencies first, then install ONNX Runtime C/C++.

Ubuntu desktop x86-64 ​

bash
sudo apt update
sudo apt install -y build-essential cmake wget tar

ORT_VERSION=1.26.0
ORT_ARCH=x64
wget https://github.com/microsoft/onnxruntime/releases/download/v${ORT_VERSION}/onnxruntime-linux-${ORT_ARCH}-${ORT_VERSION}.tgz
tar xf onnxruntime-linux-${ORT_ARCH}-${ORT_VERSION}.tgz
sudo rm -rf /usr/local/include/onnxruntime
sudo mkdir -p /usr/local/include/onnxruntime
sudo cp -r onnxruntime-linux-${ORT_ARCH}-${ORT_VERSION}/include/* /usr/local/include/onnxruntime/
sudo cp onnxruntime-linux-${ORT_ARCH}-${ORT_VERSION}/lib/libonnxruntime.so* /usr/local/lib/
sudo ldconfig

Raspberry Pi 64-bit OS ​

Use the aarch64 ONNX Runtime package on Raspberry Pi OS 64-bit or Ubuntu Server 64-bit.

bash
sudo apt update
sudo apt install -y build-essential cmake wget tar

ORT_VERSION=1.26.0
ORT_ARCH=aarch64
wget https://github.com/microsoft/onnxruntime/releases/download/v${ORT_VERSION}/onnxruntime-linux-${ORT_ARCH}-${ORT_VERSION}.tgz
tar xf onnxruntime-linux-${ORT_ARCH}-${ORT_VERSION}.tgz
sudo rm -rf /usr/local/include/onnxruntime
sudo mkdir -p /usr/local/include/onnxruntime
sudo cp -r onnxruntime-linux-${ORT_ARCH}-${ORT_VERSION}/include/* /usr/local/include/onnxruntime/
sudo cp onnxruntime-linux-${ORT_ARCH}-${ORT_VERSION}/lib/libonnxruntime.so* /usr/local/lib/
sudo ldconfig

If your Pi uses a 32-bit OS, install a 64-bit OS first or build ONNX Runtime from source for that target.

Build OpenKAI ​

Enable OpenCV, detectors, vision, and ONNX Runtime:

bash
mkdir -p build
cd build
cmake .. \
  -DUSE_OPENCV=ON \
  -DUSE_ONNXRUNTIME=ON \
  -DWITH_DETECTOR=ON \
  -DWITH_VISION=ON \
  -DONNXRuntime_include=/usr/local/include/onnxruntime \
  -DONNXRuntime_lib=/usr/local/lib
make -j$(nproc)

Export A YOLO26 ONNX Model ​

Download one of the COCO detection .pt weights:

bash
mkdir -p ~/dev/models
cd ~/dev/models

wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26n.pt
wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26s.pt
wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26m.pt
wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26l.pt
wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26x.pt

For Raspberry Pi, start with yolo26n.pt or yolo26s.pt.

bash
sudo apt install -y python3-full python3-venv

python3 -m venv ~/venv/ultralytics
source ~/venv/ultralytics/bin/activate
python3 -m pip install -U pip
python3 -m pip install -U ultralytics

python3 - <<'PY'
from ultralytics import YOLO
model = YOLO("/home/kai/dev/models/yolo26n.pt")
model.export(format="onnx", imgsz=640)
PY

Ubuntu 24.04 and newer protect the system Python environment. If pip install shows externally-managed-environment, use the virtual environment commands above instead of --break-system-packages.

The default YOLO26 export uses end-to-end detection, so _YOLO26detectONNX filters detections by confidence and does not run external NMS. If you export with end2end=False, _YOLO26detectONNX will parse the traditional output and apply OpenCV NMS using score and nms.

Config Example ​

Declare camRGB as { "type": "dataObject", "class": "RGBframe" } and set the camera's RGBframeOut output to camRGB. Declare a BBoxStream DataObject for BBoxStreamOut; consumers such as _APmav_follow read it through BBoxStreamIn. The detector keeps a local image copy while inference runs and appends detected objects to the stream's bounded history. Empty results preserve existing history; consumers filter each object's timestamp to discard stale targets. Boxes use pixel coordinates and capture timestamps. Container dimensions are separate stream metadata and assume a consistent image source for the retained history; see DataObject interfaces.

json
{
  "camRGB": { "type": "dataObject", "class": "RGBframe" },
  "detections": { "type": "dataObject", "class": "BBoxStream" },
  "YOLO26detectONNX": {
    "class": "_YOLO26detectONNX",
    "bON": true,
    "thread": { "FPS": 30 },
    "bLog": true,
    "RGBframeIn": "camRGB",
    "BBoxStreamOut": "detections",
    "fModel": "/home/kai/dev/models/yolo26n.onnx",
    "confidence": 0.25,
    "score": 0.45,
    "nms": 0.5,
    "bLetterBoxForSquare": true,
    "vModelInputSize": [640, 640],
    "bSwapRB": true,
    "scale": 0.003921568627,
    "nThread": 1
  }
}

Notes ​

  • confidence is used for the default YOLO26 end-to-end output.
  • score and nms are used for one-to-many exports with external NMS.
  • vModelInputSize must match the ONNX export image size unless the model was exported with dynamic input shapes.
  • On Raspberry Pi, start with yolo26n.onnx and keep nThread low; increasing it can help throughput on some boards but can also starve the camera pipeline.

References: Ultralytics YOLO26, Ultralytics end-to-end detection guide, ONNX Runtime install.

OpenKAI project documentation