Wednesday, 22 March 2017

Adobe’s New Algorithm Makes Super Difficult Selections a Cinch

Adobe Research has been working on some pretty interesting photo and video editing features, but their latest creation might just revolutionize photo editing as we know it. Imagine if Photoshop could automatically cut out complex subjects in seconds, no matter the background… no pen tool required.

Adobe’s selection tools have caught a lot of flack—especially the new Select and Mask environment—but this research intends to silence all the naysayers. Using a two-part process that involves a “deep convolutional encoder-decoder network” and a “small convolutional network,” Adobe can cut out any subject from in front of any background with ease and accuracy.

Keep in mind, this is 100% software-driven; there are no dual camera systems, green screens, or manual selection tricks required to make it work. And it works like gangbusters! In the example below, the original image is on the far left, the second and third images use competing methods, and the far right image is the mask created by Adobe’s software:

The accuracy demonstrated here and in latter examples in the full research paper is impressive to say the very least. Fine hair is cut out of complex and similarly-colored backgrounds with insane accuracy, and Adobe’s system seems to beat every other competing approach hands down.

The research team applied their approach to 1,000 test images after training their neural nets with a data set of 49,300 images, and the algorithm seems to have learned well.

While we have no idea when this research will make its way into a future version of Photoshop or After Effects, the impact of highly-accurate automatic selection and masking tool is not to be underestimated. Read the full paper for yourself at this link, and keep an eye out for some incredible advances from Adobe in the coming months and years.

(via DIYP)



from PetaPixel https://petapixel.com/2017/03/22/adobes-new-algorithm-makes-super-difficult-selections-cinch/

No comments:

Post a Comment