Does Frigate+ use downscaled images for training?

[nuansed]

Hi, I've uploaded a few 3840x2160 pixel images to Frigate+, but they all seem to have been downsized to a 640x360 (also looks like a nearest neighbour algo caused lots of artefacts). Is the downsized image just a low res preview to reduce bandwidth usage while the full-sized image is utilised in the backend?

Example
Cropped original image:

Cropped image from https://plus.frigate.video/dashboard/edit-image/?id=...

[NickM-27]

Yes you've got it, annotations are served on a lower res version to conserve bandwidth, the actual training is done on higher res image

[nuansed]

Ah ok, good to know. It'd be helpful to have less artefacts, maybe with a bilinear resize.

It's also a bit difficult to accurately set the bounding box on the small subject above. How about resizing images to 30% of their original size instead? Maybe with a minimum of 640x360?