Basically it's a complicated academic process that tries to automatically figure out where non-rectangular chunks of an image can be automatically identified and removed to resize the image to a target size while preserving the "important parts" of the image.
Look, this is a wonderful academic subject, but it falls firmly into the designation of things that can be done better, faster, cleaner, more reliably, and more elegantly with a much simpler process.
As fun as this is, you'll never see this technology in real use -- it's just a clever hack that's not going to yield good results on normal images in a practical way.
Trying to automate this is going to result in some very ugly images with very obvious seams and unpleasant artifacts.
Sometimes the simpler approach is the better approach.
If you want to be smart about resizing images to fit them into different dimensions, the answer is very simple:
- Create an image format that allows the author to easily and optionally indicate multiple preferred crop regions or multiple actual different images of different sizes and have a server or client display script that chooses the best crop regions to use given the available display space, with some minimal resizing to do final tweaks.
Such an approach would also let the client zoom or move around the image if they wanted. Folks, this isn't rocket science, you're over-complicating things when a simpler approach would work better and could be implemented tomorrow.