Modify the Detection Model#

How to run detection with a model other than the bundled YOLO11s — another Ultralytics size, or your own fine-tuned detector — and point one or more cameras at it. See Manage the Detection Model for the bundled model and how it is produced.

What the pipelines accept#

gvadetect loads an OpenVINO™ IR model (.xml + .bin). The model file must live below resources/models/ (the pipeline schema rejects any other path), and a camera refers to it by its path relative to that folder. PyTorch or ONNX models must be converted first.

Option A — Another Ultralytics model#

Use the virtual environment make model created to export any Ultralytics detection model (yolo11n, yolo11m, yolov8s, …) with the same options the bundled export uses:

mkdir -p resources/models/yolo11m && cd resources/models/yolo11m
../../venv/bin/yolo export model=yolo11m.pt format=openvino dynamic=True opset=18 imgsz=640 half=True
cd -

This downloads yolo11m.pt and writes resources/models/yolo11m/yolo11m_openvino_model/yolo11m.xml. Keep the pinned ultralytics==8.4.67; newer versions export a graph that fails on GPU and NPU (see Manage the Detection Model).

Option B — A custom or fine-tuned model#

  1. Convert to OpenVINO™ IR if needed:

    resources/venv/bin/yolo export model=/path/to/best.pt format=openvino dynamic=True opset=18 imgsz=640 half=True   # Ultralytics-trained
    # or, for other frameworks:
    resources/venv/bin/pip install "openvino>=2026.0" && resources/venv/bin/ovc /path/to/model.onnx --compress_to_fp16
    
  2. Copy the .xml + .bin pair into a folder below resources/models/, e.g., resources/models/my-detector/model.xml.

Models exported from Ultralytics carry their class names and post-processing in the IR’s metadata, which gvadetect reads directly. A model from another framework may need a model-proc JSON for its output layout; see the DL Streamer documentation for gvadetect model-proc=.

Point a camera at the model#

Set detection.model for the camera in configs/cameras-pymavlink.yaml (or cameras-uavsdk.yaml), relative to resources/models:

cameras:
  - id: front_camera
    detection:
      device: gpu
      model: yolo11m/yolo11m_openvino_model/yolo11m.xml
make check-cameras    # fails if the file does not exist
make start-rtsp       # or disarm/arm; the file is re-read at every arming

Cameras may use different models; each device/model pair loads its own model instance. The file is validated before anything starts, and a missing model is reported as cameras[front_camera]: detection model not found: ....

For a pipeline started by hand, pass the container path in parameters.detection_model (see Manage Pipelines — Start a pipeline):

"parameters": {"detection_model": "/home/pipeline-server/resources/models/yolo11m/yolo11m_openvino_model/yolo11m.xml"}

Verify#

make cameras-status
ffplay -rtsp_transport tcp rtsp://<HOST_IP>:8555/uav-front-camera

Boxes and labels should reflect the new model’s classes. The detection log (destination.metadata, see View Detection Logs) shows the labels as text.

Troubleshooting#

No detections with the new model (threshold), 400 from the REST API (path outside resources/models/), a model that fails on GPU/NPU, and detection model not found are covered in Troubleshooting — Models & Captioning. A larger model may not keep up with 30 fps on the CPU — use the GPU or NPU, or raise inference-interval in the pipeline string.