
YouTube goes into detail on how it is making this all happen, and as you would expect it’s an interesting process that was developed during the initial stages of the beta. First they analyzed the color, spatial layout, and motion of regular 2D video to create a depth map for each individual frame. Then they shifted that information over to another process where the map could be compared with video that was shot in true 3D. Using some nifty machine learning, an automated process applies the depth estimation that it gleaned from the previous analysis. This results in the creation of a stereoscopic video from the original “flat” source as illustrated below.
After all that was accomplished, it was just a matter of scaling the process up to be able to convert a larger number of videos that are stored on YouTube’s servers. Having tried this out on a few videos using a simple pair of 3D glasses my kids have stashed around the house, I can say that the conversion process works pretty well.
The video at the top of the post is the best I have seen out of the few that I previewed, but I’m sure that over time the selection will improve.
Read more at YouTube’s official blog