Manage the Detection Model#

The application detects objects with YOLO11s, Ultralytics’ COCO-pretrained small model, exported to OpenVINO™ IR by make model. This page covers the bundled model, how it is produced, where it lives, how the pipelines find it, and how to re-export it. To use a different model, see Modify the Detection Model; for the scene-captioning model, see Change the Captioning Model.

The bundled model#

Property

Value

Model

YOLO11s (Ultralytics, COCO-pretrained)

Format

OpenVINO™ IR, FP16, dynamic shape, opset 18, 640 × 640 input

Classes

80 COCO classes (person, car, truck, bus, bicycle, motorcycle, …)

Export tool

ultralytics==8.4.67 + openvino>=2026.0 (resources/requirements.txt)

Important

ultralytics is pinned to 8.4.67. Since release 8.4.115, releases export a detection head that compiles on CPU only and fails on the GPU and NPU plugins. Do not bump the pin without re-verifying the exported model on all three devices.

Download and export (make model)#

Prerequisites: Python 3.10+ with python3-venv (sudo apt install python3-venv on Ubuntu 24.04 and 26.04), and internet access (set https_proxy/http_proxy first behind a proxy).

make model

The target first shows the model’s license and asks for confirmation (ACCEPT_LICENSE=y skips the prompt, for scripts). It then creates resources/venv/, installs resources/requirements.txt, downloads yolo11s.pt and runs yolo export format=openvino half=True dynamic=True imgsz=640. Output:

resources/models/yolo11s/
├── yolo11s.pt                       # downloaded checkpoint
└── yolo11s_openvino_model/
    ├── yolo11s.xml                  # OpenVINO IR
    ├── yolo11s.bin
    └── metadata.yaml

make pymav-up / make uavsdk-up refuse to start until yolo11s.xml exists.

How the pipelines find the model#

resources/ is mounted read-only in the container at /home/pipeline-server/resources/. Each camera names its model in configs/cameras-pymavlink.yaml relative to resources/models:

    detection:
      device: npu
      model: yolo11s/yolo11s_openvino_model/yolo11s.xml

The pipeline manager turns this into the pipeline’s detection_model parameter (/home/pipeline-server/resources/models/<model>) and a detection_instance_id derived from device and model file, so cameras that share a device and model share one loaded model. The pipelines’ own default for detection_model is the bundled YOLO11s path.

Re-export#

Delete the export and run make model again, for example after changing resources/requirements.txt or to try another export option:

rm -rf resources/models/yolo11s/yolo11s_openvino_model
make model

For an INT8 export add int8=True to the yolo export call in the Makefile’s model target (nncf is already in the requirements). Restart the stack afterwards so running pipelines load the new files.

Verify#

ls resources/models/yolo11s/yolo11s_openvino_model/
make check-cameras         # reports a camera whose model file is missing

Troubleshooting#

make model failing on python3-venv, a model that loads on CPU but not on GPU/NPU, and a missing model reported by make check-cameras are covered in Troubleshooting — Setup and Troubleshooting — Models & Captioning.