I do street shooting a few hours a day, with a backlog of 300K pics I want to handle with quick accept/reject clicks in a GUI that gives me crop and rotation suggestions, and ideally auto color editing based on the 100K pics I’ve edited manually. The first two I believe are simple enough for a PC with existing neural net tech, while color editing seems like it might need some 240-volt GPU power.
I want a camera-to-website editing pipeline where, as I look at the pics, I optionally see/select the crops and have a ‘straighten this’ button available. Then try color edit suggestions, save, and the photo moves on to content analysis, which is what I’ve been working on.
I’m using the ‘parsing’ of popular classification neural networks to train simple nets to match interesting pairs of photos - ‘Rorschach pairs’, which can be done with >90% accuracy for practical purposes by using the top predictions. The rankings of pairability also give an interesting way to find some of the better photos in a big pile, and show how training a net to rank your photos directly may be straightforward.