The web quickly embraced DeepDream and its surreal images. A huge number of users mercilessly squeezed the "imagination" of neural networks, yet we shouldn't forget that those images were created with even more complex functions in mind. A good proof of this is DeepStereo, another Mountain View system capable of developing animations from a series of images, calculating and filling the "digital voids" with more than acceptable precision.
With ten or twenty images in hand, the user can't go much beyond the classic slideshow to present them animatedly. Video is an excellent alternative, but unless it undergoes a good amount of post-processing, the result filmed with a mobile device will be unstable and chaotic. Now, there is the possibility of obtaining more information from those images. Traditional tools fall by the wayside, but when analyzed with the assistance of a neural network, it's possible to produce data that isn't there. From a certain point of view, that's how DeepDream works, whose images contain animals and structures out of nowhere. DeepStereo is even more robust.
Instead of drowning the system in multiple processing phases and custom adjustments, DeepStereo was trained by Google end-to-end with a specific objective: when several images are loaded, the system extracts pixels from different angles and, from that information, generates the pixels that aren't observed in the first place. The result of moving through the images is an animation with complete transitions and minimal errors, detected mainly in lighting differences. DeepStereo also takes into account color and depth details, so a lamp doesn't end up trapped inside a column, or embedded in the wall.
Needless to say, DeepStereo's work isn't perfect. If the difference in time and lighting between images is too large, errors arise during the transformation process, but that doesn't erode its enormous potential. Imagine a Street View version with smooth transitions, or why not, cinematic and virtual reality projects backed by this technology. As with DeepDream, it wouldn't be bad for Google to release DeepStereo's code, but something tells me we'll have to wait a bit...
