Hello I'm running some experiments with yolov8n-segmentation. I have successfully convert and deploy my weights in a a RVC2 Luxonis camera. However something that I've notice while conducting the experiments is that the performance from my model is downgraded on the Luxonis device. I've run the ONNX weights on a CPU and it performs better than the .BLOB weights on the Luxonis size.

I convert the model using the https://blobconverter.luxonis.com/ tool with mean = 0 and scale = 255. Any idea how to improve the performance so in can match the ONNX?

    AlejandroDiaz
    Well this depends on the complexity of the model. If you have a performant CPU, it can rival the RVC2 VPU (1.4 TOPS).

    Thanks,
    Jaka

    My model is the yolov8n-seg and my CPU is a regular Rasbberry PI 4B. In terms of inference speed the luxonis is faster however in terms on the quality of the segmentation masks, the ONNX running on the CPU performs much better than the BLOB running on the Luxonis camera. Is there a tool like netron that allow me to see the architecture of my weights?