Some questions about model conversion

[user21]

I apologize for asking some basic questions. I have trained several YOLOv8 models and successfully deployed them on a Hailo-8 + Raspberry Pi 5 setup. However, simply converting the official models isn’t enough for my use case—I work in pest detection, where the targets are very small. The official YOLO models perform reasonably well for small objects, but once converted to .hef format and run on the Raspberry Pi, the detection accuracy drops significantly.

I’m wondering if adding a small-object detection head or attention mechanisms could improve accuracy. I know how to modify the network structure in the official YOLOv8 code, but I’m unsure what corresponding changes need to be made to the Hailo translation framework. Could you point me to any relevant documentation for this?

Additionally, if I want to extract the four bounding box coordinates (within a video frame), where can I find documentation or guidance on how to implement this?

[nina-vilela]

Hey @user21,

@user21

The official YOLO models perform reasonably well for small objects, but once converted to .hef format and run on the Raspberry Pi, the detection accuracy drops significantly.

If you are having accuracy issues, we recommend going through this section of our guide.

@user21

I’m unsure what corresponding changes need to be made to the Hailo translation framework. Could you point me to any relevant documentation for this?

The flow would be pretty much the same. Depending on your changes you might have to change the NMS config json.

@user21

Additionally, if I want to extract the four bounding box coordinates (within a video frame), where can I find documentation or guidance on how to implement this?

We have this example in c++, and this in Python.