Slow inference with yolov5s.onnx vs onnxruntime Mac M1

[Christof23]

Hi, really excited about this project. Have been looking into using tract for object detection inference on CPU and can get around 73-75ms inference speed with onnxruntime but only around 300ms with tract on Mac M1. Am looking to export for a Raspberry Pi 4 ARMv8 target, so hoping for some improved relative performance, but might there be anything I'm missing for apple CPU? Or is Mac M1 generally not the best place for tract? Would be fantastic to get similar performance to remove these non-rust libs!