Two questions:
1: How does the hailo picamera2 library create the low res image for inference? Does it crop it or stretch it?
2: If I plan to train a model on a specific resolution, does it make sense to have all my captured detection images that I will be using for training at that same resolution? My logic is that there is a constant resolution through the whole pipeline, I understand it’s good to have variation to make a model more robust but I plan on keeping this camera pointing at the same spot forever.
Since Picamera2 is specifically a Raspberry Pi application, you’ll probably get better support by asking in their community directly.
Regarding model training - you’re absolutely right that using the original resolution is ideal when possible. However, there are workarounds if you need them. For example, you can use a tiling approach where you split a large image into smaller tiles, run detection on each tile, and then piece together the results to get inference on the full image. This really depends on your specific application and model requirements though.
From what I can see in your case, it looks like you should be able to run at the exact resolution you need.