Facebook announcing the completion of an internal research project dedicated to helping you clean up poorly taken 360 photos. The company laid out a system that uses deep neural networks to try to correct for common orientation errors with the photos that upload.
If user taking a 3600 photo doesn’t hold the camera perfectly in line. The resulting image can tilt, which makes it harder to read and broken the sense of immersion if the image viewed in virtual reality.
Facebook’s system takes on a photo and outputs a pair of values for the tilt and roll correction needed to bring the horizon of the photo in line. It’s based on AlexNet, an image recognition system that used to tackle other problems like determining the contents of the images.
Making 3600 photos look good on Facebook is key for the social networking company. The company invests heavily in virtual reality. In addition to the automatic rotation issue, Facebook also had to contend with the massive size of the 3600 photos that uploaded to its service. While that may not be a huge problem with super-fast networks and devices, it could be an issue on mobile devices on cellular networks.
Facebook converts the photos into cubes, and stores those cubes at different resolutions. Those images broken up into a set of 512×512-pixel squares. When a user pulls up a photo, Facebook calculates resolution and position within the image needs to load.
In the event it’s not possible to get a high enough resolution right away, the social network provides a lower-res version until the correct quality is available.