Hi, I’m a beginner at developing edge models with Raspberry Pi 5. I converted my MiDaS ONNX model to a .hef file for deployment. I followed the dataflow guide on my own, but I think my model is quite larger than the model zoo depth example. Is there more aggressive quantization I can apply to achieve 30fps on the Hailo8 chip?
Here’s a comparison of the output: FP32 ONNX (upper) vs INT8 quantized (below).
HEF file size is about 20MB
