Completely foreign field to me, and I only skimmed the text, but argumentation against better methods vs cropping and downsampling sounds surprising as we have had image compression methods that try to preserve detail for a long time? Is it that this is nicer for applications where the image is intended as input for computer vision algorithms? Maybe this is computationally simpler so there's not as much overhead in accessing the photo details vs a compressed image?
One of the authors here! Happy to answer any questions people have (w/ Ritwik)