Optimizing 360 photos at scale

This post details the re-engineering of Facebook's photo infrastructure to support high-resolution 360 photos through tiled storage and serving. It also describes the application of a modified AlexNet deep neural network to automatically correct the tilt and roll of 360 photos, improving realism. The DNN was trained on monochrome equirectangular images and cast as a regression problem to predict tilt and roll values, achieving an average accuracy of 0.1 degrees.