I kick myself every time I mess up the settings on my camera. On Nicole’s first day of school I mistakenly left my camera in full manual mode, with the shutter speed set for indoor photography of stationary objects. The photo on the left is my favorite pose wise. Nicole is so happy. But the photo is so over exposed with a slight motion blur given the slow shutter speed. Fortunately I realized my mistake, adjusted my settings and was able to also capture the photo on the right.

Current state of the art photo editing techniques cannot save the photo on the left. The skin on the left side of her face is so over exposed the sRGB value is pure white. The computer simply has no way to figure out what color was supposed to be there, so there’s no way to automatically fix the white patch. Nicole could have cyan colored skin for all the computer knows. Or checker pattern skin, for that matter.

As an machine learning person, I find this limitation incredibly irritating. Chances are, when you, a human, first looked at the image on the left Nicole’s skin did not appear pure white to you. Your brain filled in a likely color based on the right side of the photo where her skin isn’t as badly over exposed and what you know of human skin tones. With the two photos side by side you can easily imagine a combined photo with the pose from the left and the coloring on the right.

Current state of the art image recognition can recognize faces. It should be able to map point for point the location of the eyes, mouth, nose, hands, etc between the two photos. Given that, it seems plausible for a AI enabled photo editing software to merge the two photos and create the ideal image.

Even without the image on the right, a sufficiently advanced AI based system should be able to recognize the subject matter much the way a human would. The system should then be able to generate a hyper realistic plausible image. It might not be the “correct” image, the image that would have been created if my settings were correct in the first place. I doubt for most people that would matter. A realistic enough resulting image that was plausibly correct would likely be sufficient for most momtographers like me.

Posted in Photography | Tags: Momtographer, Photo Editing