Tuesday, April 19, 2011

Microsoft Research’s AutoCollage

Those who may have read my blog previously will know I like making photo collages, often in picasa but I’m open to any good collage tool. Autocollage works are something I first saw on Flickr, and didn’t realise at the time it was another microsoft reseach project (PhotoSynth & ICE are two others I have discussed before). These collages at first appear like very professional graphic artistry and beautifully “seamless” but they are really cleverly blended image segments, that any of us can create with relatively little effort (ok you need to take some good photos first).

It works best with photos of family and friends, say on special occasions. It has been built around a robust multi-pose face recognition system (yes it does sound a bit big brother-ish). The Collage construction is automatic and first looks for the most interesting photos, and particularly clear face shots and the regions of interest (thats ROI Optimization to researchers) with a constraint satisfaction optimization (lost you yet, it means it is trying to pack as many ROI into the collage space as efficiently as possible. A really smart cut-paste system called grab-cut (or as it is know to researches, Image Segmentation with A Bounding Box Prior) then selects the key parts of those ROI and finally a neat blender algorithm that solves a linear system of Poisson Equation to get seamless pastes the segments into the full collage image. Simple really! Well it is to use.

New folder_AutoCollage_20_Images_4
I make a point of not posting links to “hostage”ware (software you can download and use once or twice but then it demands payment) but this download is clearly a 30 day free trial. You do however get a big watermark across the bottom of each image generated with the trial version.

No comments: