README.MD

Object detection - yolo26 with lightweight ONNX postprocessing

This example demonstrates object detection using a Hailo-8 device, based on the 2026 Ultralytics YOLO26 release of top-performing NMS-free networks and a split HEF + ONNX postprocessing pipeline.

It processes input images, videos, or a camera stream and annotates them with the detected objects. Optionally, object tracking across frames can be enabled for video and camera streams.

ONNX postprocessing

Similarly to hailo_apps/cpp/onnxrt_hailo_pipeline, this example uses onnxruntime for the postprocessing part. This makes integration of new networks especially convenient, by following these steps:

1. Split the ONNX into the "neural processing" and the "postprocessing" parts using extract_postprocessing.py script
2. Process the first part into a HEF using the DFC
3. In runtime, apply the second part on the HEF outputs with onnx-runtime engine to complete an accelerated equivalent of the original ONNX. This runtime part is implemented and exemplified in this app.
4. The desired "full-onnx = HEF + postproc-onnx" equivalence can be conveniently debugged (isolating pipeline vs. compilation issues and HEF degradation) using --full-onnx flag that applies a bypass of the HEF (using the 'neural-processing' split-onnx 1st part). This is also useful for quick 'dry' tests without hardware or compilation at all - as well as benchmarking the acceleration provided by Hailo's offloading of the neural part to HEF running on hardware.

Requirements

• hailo_platform:
  • 4.23.0 (for Hailo-8 devices)
  • 5.3.0 (for Hailo-10H devices)
• opencv-python
• scipy
• lap
• cython_bbox
• onnxruntime
• python-dotenv
• PyYAML

Supported Models

• YOLOv26

Linux Installation

Run this app in one of two ways:

1. Standalone installation in a clean virtual environment (no TAPPAS required) — see Option 1
2. From an installed hailo-apps repository — see Option 2

Option 1: Standalone Installation

To avoid compatibility issues, it's recommended to use a clean virtual environment.

0. Install PCIe driver and PyHailoRT

   • Download and install the PCIe driver and PyHailoRT from the Hailo website
   • To install the PyHailoRT whl:

   pip install hailort-X.X.X-cpXX-cpXX-linux_x86_64.whl

1. Clone the repository:

   git clone https://github.com/hailo-ai/hailo-apps.git
   cd hailo-apps/python/standalone_apps/object_detection_onnx_postproc

2. Install dependencies:

   pip install -r requirements.txt

Option 2: Inside an Installed hailo-apps Repository

If you installed the full repository:

git clone https://github.com/hailo-ai/hailo-apps.git
cd hailo-apps
sudo ./install.sh
source setup_env.sh

Then the app is already ready for usage:

cd hailo-apps/python/standalone_apps/object_detection_onnx_postproc

Run

After completing either installation option, run from the application folder:

./object_detection_onnx_postproc.py -n <model_path> -i <input_path>

Windows Installation

To avoid compatibility issues, it's recommended to use a clean virtual environment.

0. Install HailoRT (MSI) + PyHailoRT

   1. Download and install the HailoRT Windows MSI from the Hailo website.

   2. During the installation, make sure PyHailoRT is selected (in the MSI “Custom Setup” tree).

   3. After installation, the PyHailoRT wheel is located under: C:\Program Files\HailoRT\python

   4. Create and activate a virtual environment:

   python -m venv wind_venv
   .\wind_venv\Scripts\Activate.ps1

   5. Install the PyHailoRT wheel from the MSI installation folder:

   pip install "C:\Program Files\HailoRT\python\hailort-*.whl"

1. Clone the repository:

   git clone https://github.com/hailo-ai/hailo-apps.git
   cd hailo-apps\hailo_apps\python\standalone_apps\object_detection

2. Install dependencies:

   pip install -r requirements.txt
   
   

Run

After completing either installation option, run from the application folder:

python .\object_detection_onnx_postproc.py -n <model_path> --onnx <onnx_path> --onnx-config <onnx_config_path> -i <input_path> 

You can choose between:

• Object detection
• Object detection with tracking (add --track)

The output results will be saved under a folder named output, or in the directory specified by --output-dir.

Arguments

• --hef-path, -n:
  • A model name (e.g., yolov8n) → the script will automatically download and resolve the correct HEF for your device.
  • A file path to a local HEF → the script will use the specified network directly.
• -i, --input:
  • An input source such as an image (bus.jpg), a video (video.mp4), a directory of images, or usb to auto-select the first available USB camera.
    • On Linux, you can also use /dev/vidoeX (e.g., /dev/video0) to select a specific camera.
    • On Windows, you can also use a camera index (0, 1, 2, ...) to select a specific camera.
    • On Raspberry Pi, you can also use rpi to enable the Raspberry Pi camera.
  • A predefined input name from resources_config.yaml (e.g., bus, street).
    • If you choose a predefined name, the input will be automatically downloaded if it doesn't already exist.
    • Use --list-inputs to display all available predefined inputs.
• -b, --batch-size: [optional] Number of images in one batch. Defaults to 1.
• -l, --labels: [optional] Path to a text file containing class labels. If not provided, default COCO labels are used.
• -s, --save-output: [optional] Save the output of the inference from a stream.
• -o, --output-dir: [optional] Directory where output images/videos will be saved.
• --draw-trail: [optional][Tracking only] Draw motion trails of tracked objects.
• cr, --camera-resolution: [optional][Camera only] Input resolution: sd (640x480), hd (1280x720), or fhd (1920x1080).
• or, --output-resolution: [optional] Set output size using sd|hd|fhd, or pass custom width/height (e.g., --output-resolution 1920 1080).
• --track: [optional] Enable object tracking across frames using BYTETracker.
• --show-fps: [optional] Display FPS performance metrics for video/camera input.
• --no-display: [optional] Run without opening a display window. Useful for headless or performance testing.
• --video-unpaced: [optional] Process video input as fast as possible without respecting the original video FPS (no pacing).
• -t, --time-to-run: [optional] Maximum runtime in seconds. Stops the application after the specified duration.
• -f, --frame-rate: [optional][Camera only] Override the camera input framerate.
• --list-models: [optional] Print all supported models for this application (from resources_config.yaml) and exit.
• --list-inputs: [optional] Print the available predefined input resources (images/videos) defined in resources_config.yaml for this application, then exit.
• --onnx ONNX_PP_FILE: [optional] Override path to ONNX postprocessing model file (2nd part of split). If omitted, use existing resource lazy-downloaded from preconfigured cloud path (alongside the HEF)
• --onnx-config ONNX_CONFIG_FILE: [optional] Path to the ONNX postprocessing configuration file. If omitted, a default configuration is used if available.

For more information:

./object_detection_onnx_postproc.py -h

Example

List supported networks

./object_detection_onnx_postproc.py --list-nets

List available input resources

./object_detection_onnx_postproc.py --list-inputs

Inference on a usb camera stream

./object_detection_onnx_postproc.py -i usb

Inference with tracking on a usb camera stream

./object_detection_onnx_postproc.py -i usb --track

Inference with tracking and motion trail visualization

./object_detection_onnx_postproc.py  -i usb --track --draw-trail

Inference on a usb camera stream with custom frame rate

./object_detection_onnx_postproc.py -i usb -f 20

Inference on a video

./object_detection_onnx_postproc.py  -i full_mov_slow.mp4

Inference on an image

./object_detection_onnx_postproc.py -i bus.jpg

Inference on a folder of images

./object_detection_onnx_postproc.py -i input_folder

🔧 Visualization and Tracking Configuration

The application supports flexible configuration for how detections and tracking results are visualized. These settings can be modified in the configuration file to adjust the appearance of detection outputs and the behavior of the object tracker.

Example Configuration:

"visualization_params": {
    "score_thres": 0.42,
    "max_boxes_to_draw": 30,
    "tracker": {
        "track_thresh": 0.01,
        "track_buffer": 30,
        "match_thresh": 0.9,
        "aspect_ratio_thresh": 2.0,
        "min_box_area": 500,
        "mot20": false
    }
}

Parameter Descriptions:

Visualization Parameters:

• score_thres: Minimum confidence score required to display a detected object.
• max_boxes_to_draw: Maximum number of detected objects to display per frame.

Tracker Parameters:

• track_thresh: Minimum score for a detection to be considered for tracking.
• track_buffer: Number of frames to retain lost tracks before deleting them.
• match_thresh: IoU threshold used to associate detections with existing tracks.
• aspect_ratio_thresh: Maximum allowed aspect ratio of detected objects (used to filter invalid boxes).
• min_box_area: Minimum area (in pixels) of a detection to be considered valid for tracking.
• mot20: Whether to use MOT20-style tracking behavior (set to false for standard tracking).

Additional Notes

• The example was tested with:
  • HailoRT v4.23.0 (for Hailo-8)
  • HailoRT v5.3.0 (for Hailo-10H)
• The example expects a HEF which contains the HailoRT Postprocess
• Images are only supported in the following formats: .jpg, .jpeg, .png or .bmp
• Number of input images should be divisible by batch_size
• The list of supported detection models is defined in networks.json.
• For any issues, open a post on the Hailo Community

Disclaimer

This code example is provided by Hailo solely on an “AS IS” basis and “with all faults”. No responsibility or liability is accepted or shall be imposed upon Hailo regarding the accuracy, merchantability, completeness or suitability of the code example. Hailo shall not have any liability or responsibility for errors or omissions in, or any business decisions made by you in reliance on this code example or any part of it. If an error occurs when running this example, please open a ticket in the "Issues" tab.

This example was tested on specific versions and we can only guarantee the expected results using the exact version mentioned above on the exact environment. The example might work for other versions, other environment or other HEF file, but there is no guarantee that it will.