Facebook announced today new software techniques for encoding and streaming video including video from 360-degree cameras and for virtual reality goggles. Clearly the Web giant wants to drive more rich content to its data centers, whether or not you think the Oculus Rift and its competitors will be big hits or belly flops this year.
The move is typical of Facebook which wants to get an advantage by driving all development work into the open. Its larger rivals Google and Apple tend to do their work behind closed doors, so we can assume they have similar projects in their labs. By giving away more goodies, Facebook hopes to attract partners and gain needed leverage.
Facebook described in a blog and an event at its headquarters today a handful of projects. For example, it released code on GitHub for a custom filter to transform user-uploaded 360 videos into a cube map format. The format reduces files sizes by 25%, Facebook said.
While it didn’t release any code specific to VR, Facebook did describe its latest techniques. For example, it talked about using pyramid geometry to reduces file size by 80 percent when encoding 360-degree video for VR. It also described an approach for view-dependent adaptive bitrate streaming for optimizing a VR experience.
Imagine someday putting on your VR goggles to view and share updates on Facebook! I imagine you are either ready to jump for joy – or puke.
Driving all that rich content through its servers will surely have a cost. Facebook described early work on an artificial intelligence scheme for mining data from that video that could be part of how it makes its ROI.
Specifically, Facebook’s Vision Understanding team showed “a novel AI architecture that is an early step toward unsupervised understanding of what is happening in a video and predicting what's going to happen next… Instead of labeling objects, scenes, and actions, these networks label voxels, or individual video pixels over time.”
I suspect we will eventually hear privacy debates about whether or not we want our personal voxels labeled. Do you need to own your video metadata?
Incidentally, while doing all this fancy 360-degree video and VR work, Facebook also built a new streaming video engine (SVE) for transcoding video. It delivers a 10x improvement in the processing time between video upload and playback, Facebook claims. The Web giant migrated all of its video traffic from the old system without any interruption to its service while it was growing to serve up an astounding 8 billion video views a day.
“Rather than treat an entire video as a single file, SVE splits the video into segments…allowing parallel processing of videos and results in much lower latency,” it said.
When you have hundreds of thousands of servers, you do all the work in software. No need for special codec chips in hyper-scale data centers.
This reality should make semiconductor companies who sit next door to Facebook and Google in Silicon Valley somewhat uncomfortable. It’s not just a matter of Intel and the x86 winning all the great sockets. It’s an issue of Web giants offering free services and saying there really are no other sockets needed.
There’s a brainpower issue, too. I had heard the chief scientist of Nvidia complain he can’t hire great AI developers because they are already working for Facebook or Google. Sounds like the Web giants also have a lot of the brains in video processing generally.
Well, that’s enough work for now. Time to update my Facebook status.
— Rick Merritt, Silicon Valley Bureau Chief, EE Times