Long ago we made downscaling a global thing. This isn't ideal though, since downscaling is relatively cheap (especially at integer factors), and some detectors now do their own kind of downscaling or downsampling (e.g. perceptual hash detectors).
We need to come up with a better API to either allow detectors to indicate if they want a downscaled frame, or move all downscaling logic into the specific detectors that would require them. As a stretch goal, we would also want to avoid downscaling the frame twice at the same scale, but this isn't necessary for a first pass and could also just be done as a follow up.
Long ago we made downscaling a global thing. This isn't ideal though, since downscaling is relatively cheap (especially at integer factors), and some detectors now do their own kind of downscaling or downsampling (e.g. perceptual hash detectors).
We need to come up with a better API to either allow detectors to indicate if they want a downscaled frame, or move all downscaling logic into the specific detectors that would require them. As a stretch goal, we would also want to avoid downscaling the frame twice at the same scale, but this isn't necessary for a first pass and could also just be done as a follow up.