ONNX Runtime - PPE detection with YOLO-World

ONNX Runtime - PPE detection with YOLO-World

Using YOLO World you do not have to train the model. Instead, you just specify to the model the classes you want to detect. The classes to detect are "embedded: in the model and they were provided before exporting it to ONNX. If you want to change them, the easiest way is to do it via HuggingFace (opens in a new tab). Just write what you want to detect there, run a test, and export the model file. Then update the process.json file to point to your new file.


  • Pipeless: Check the installation guide.

  • Python OpenCV and NumPy packages. Install them by running:

pip install opencv-python numpy

Run the example

Create an empty Pipeless project

pipeless init my-project --template empty # Using the empty template we avoid the interactive shell
cd my-project

Feel free to change my-project by any name you want.

Download the stage folder

wget -O - https://github.com/pipeless-ai/pipeless/archive/main.tar.gz | tar -xz --strip=2 "pipeless-main/examples/yolo-world"

(Optional) If you have CUDA or TensorRT installed you can enable them at process.json

    "runtime": "onnx",
    "model_uri": "https://pipeless-public.s3.eu-west-3.amazonaws.com/yolow-l-ppe.onnx",
    "inference_params": {
        "execution_provider": "cpu"

Start Pipeless

The following command leaves Pipeless running on the current terminal. We add the stream-buffer-size to limit the number of frames that we accumulate, avoiding filling the memory. This is required because the model is relatively big and takes some time to run inference on CPU. Feel free to update the value or remove the option.

pipeless start --project-dir . --stream-buffer-size 35

Provide a stream

Open a new terminal and run:

pipeless add stream --input-uri "v4l2" --output-uri "screen" --frame-path "yolo-world"

This command assumes you have a webcam available, if you don't just change the input URI by a file (file:///absolute/path/to/your/file).