Facebook has long been one of the best places to share photos online, including 360 degree photos, and the company is reportedly working on an AI product based on neural networking that will help to patch up and optimize 360 degree photos that may not have been taken under optimal conditions or with the best equipment. At Facebook's latest @Scale developer conference, plans for a convolutional neural network that could be trained to fix up tilted or torn photos were laid bare. Essentially, this AI will be able to receive a 360 degree image that has resolution and detail issues, or has tears due to the camera tilting unevenly, and fix those up to look better and be more immersive for VR consumption.
The way that the neural network operates is pretty simple, on the surface. For photos with tilted sections, all it has to do is figure out how much and in what direction that part of the image must be tilted to bring the horizon back in line with what's in the other photos. Obviously, since this is dependent on a number of factors like shooting venue and subject, the AI will start off with just its training from the lab, and will get better over time. The other fixes are applied by breaking 360 degree images up into sections, then applying corrections and tweaking the resolution on a per-section basis to ensure that the entire image has a consistent look and quality, and that the look and quality on display are the best possible. This approach is also used to optimized file sizes and spare users network and device resources.
the AI that will be doing all of this is still in the earliest phases of production, but has reportedly been making good progress. 360 degree imagery and VR content are a big part of the social network's push into more modern mediums, so development of this system is likely a fairly high priority. Still, there is no definite date on when exactly this AI will go live for all users, or even begin a public field test; Facebook has only announced what its intention with the AI is, and that it's being worked on.