Hello everyone, I have this problem. My camera gives an image of 1920 by 1080 pixels, but my neck is trained at 640 by 640. Where can I find the correct frame conversion?
Welcome to the Hailo Community!
It depends on your use case and how the network was trained.
You can simply crop a region of the image or scale the image to the target size.
If you scale the image with a 1:1 aspect ratio (preserving proportions), you may end up with black bars on the top/bottom or left/right. Alternatively, you can stretch or squeeze the image to exactly fit the network input dimensions, though this may distort objects. If the network was trained with the same preprocessing, this approach can still work.
Another option is tiling during inference — dividing the original large image into overlapping or non-overlapping tiles of 640×640 and running detection on each tile. This can help with detecting smaller or distant objects.
The choice of preprocessing depends on your platform and framework (e.g., OpenCV, GStreamer).