I'm interested in the related problem of cropping an image to a new aspect ratio automatically.
Idle thought: Take a bunch of sample images, show them to humans, record heat maps of where the humans look. Use this data to train a GAN to generate heatmaps from images.
Feed the GAN a new image. Crop the image to keep the parts where the GAN "looks".
I took over an open studios site 10+ Years ago. The original web developer/ graphic designer asked for square thumbnails. then proceeded to crop them down about 30% before displaying on certain pages. Artists reacted by only uploading usable images in the remaining area (so the whole image filled the thumbnail). The uploaded images had the right and tops blank.
Sometimes it’s best to let the originator of the image control it.
Idle thought: Take a bunch of sample images, show them to humans, record heat maps of where the humans look. Use this data to train a GAN to generate heatmaps from images.
Feed the GAN a new image. Crop the image to keep the parts where the GAN "looks".
Thoughts? Any reason why it wouldn't work?