Validating theory about preview sizing/cropping

dizzyd

Hi all,

When using the preview output on a Camera node (v2.6), does it crop or scale the ISP output?

I’ve got my ISP scaling the sensor image down to ~720p, and I want to feed that into a neural network. The docs seem to suggest that the preview output crops the image to the requested size, rather than scaling it — which would mean losing a lot of sensor data.

If that’s the case, would the right approach be to use the video output instead and then run it through an ImageManip node to get a properly letterboxed/scaled image for the NN input?

Any guidance?

Thanks,

OskarSonc

Hey @dizzyd — ColorCamera.preview does a center-crop to the requested aspect ratio, then scales. It doesn’t letterbox, so you’ll lose some FOV if your NN input aspect ratio doesn’t match.

I’d recommend using the video (or isp) output for full-FOV, and feeding that into an ImageManip node with setResizeThumbnail(W, H) to letterbox it for your NN input.

Thanks,
Oskar