Hi folks -- study PI here. Happy to answer any questions.
We're working on a paper for submission to a research conference soon, so every viewer gives us helpful data for the system to learn in-situ and improve its performance over different kinds of Internet connections and congestion situations. (We are randomizing connections to get either our in-development algorithm or a bunch of competing ABR and congestion-control schemes so we can properly evaluate, so if you get a stall or bad quality...... we'll blame that.)
All the source code is at https://github.com/StanfordSNR/puffer and is or will be released with an open-source license. The study is led by my doctoral student Francis Yan, along with Sadjad Fouladi, Hudson Ayers, Chenzhi Zhu, and my colleague Philip Levis. The goal is to train video-streaming algorithms (bitrate selection and congestion control) online, continually, in situ, to minimize stalls and deliver the best picture quality across a diverse range of real-world Internet connections and users. This is from some of the same people who brought you Salsify (functional video compression for lower-latency real-time video: https://snr.stanford.edu/salsify), Lepton (distributed/parallel JPEG compression: https://github.com/dropbox/lepton), Mosh (functional terminal emulation for mobility: https://mosh.org), Remy (machine learning for congestion control; http://mit.edu/remy), and the Pantheon of Congestion Control (https://pantheon.stanford.edu).
There are some unusual details of the Puffer system to make experimentation easier (and, we hope, performance better): the ABR algorithm is server-side, and video is streamed continuously over a WebSocket and the client gives asynchronous feedback, instead of using DASH HTTP request/replies. The tcp_info struct (including the delivery rate estimate) from the congestion-control layer is plumbed through to the ABR, and to the neural network that tries to predict "how long will it take to send a video chunk of length y," so it's weakly cross-layer. The "transmission-time predictor" is probabilistic instead of simply producing a point estimate. We encode with libx264 and then measure the SSIM of every encoded chunk to calculate a reward for each possible stream (instead of making the assumption that more bitrate is better -- the correlation is not so good when you consider different times in the same video). And we encode 10 variants of each channel, so more than typical. And everything is done at 60 fps. Of course some of these details mean these algorithms are not going to be deployed any time soon on a production CDN or streaming service, but the hope is to demonstrate the value of some of the ideas in an academic setting for future commercial deployment. The use of WebSockets, MSE, and Opus is why this doesn't work on Safari or on iOS.
I'm outside the US but checked the box out of curiosity and it let me stream anyway. Am I corrupting your test? It really is fast and clean - very impressed!
My apologies that this question is likely tangential to the main objective of the research, but will this project be open sourcing the signal acquisition, transcoding, and delivery portion of its research as well?
This part of this project alone is good enough to improve, dare I claim disrupt, enormous monied interests in the global media landscape.
If this whole thing is legal (500 simultaneous peers), and can be reduced to a docker image and some rabbit ears, you could probably improve the lives of millions (60?) of Americans that do not have access to local stations.
Also, the claim of "not going to be deployed any time soon on a production CDN or streaming service", I would like to make a friendly wager that you are mistaken. I won't even demand favorable odds because of my disadvantaged intellect. This is impressive and valuable work.
Thank you for your kind words! The students worked really hard to get all those details right. And yes, all that stuff is at the GitHub repo (https://github.com/StanfordSNR/puffer).
Streaming it in another browser window as a part of my Mechanical Turk HIT assignment as I type, and I have to say -- I'm blown away by the incredible picture quality and buffer-free performance.
Puffer team has done a great job.
As a sidenote, I just realized how annoying cable TV ads are. The content to advertising ratio on cable is insanely bad. This 'Puffer' experience just validated my decision to stay away from cable TV.
That would still make you wait for the switch.
Switching channels on digital TV is slow due to key frames - you need to wait until one full Intra-frame of the new channel has been decoded. Here client can simply request Intra-frame on demand and receive one immediately.
Higher up a comment mentions they use websocket. I believe it uses the same connection, and in the backend it swaps you to a different feed, allowing it to be very quick. So really you only get one keyframe worth of black before the new video stream starts.
In the case of Locast, part of the Aereo ruling was that it was a for profit service. Locast is able to do this because they are not charging the end user.
This is very impressive. FYI: Delaying the buffer spinner, not showing it unless buffering has been (let's say) >300ms for instance, makes the switching feel a lot faster. If you could also get rid of the black between switching out the media source buffer, they would feel instant.
I'm going to be experimenting with the source code, I really need some good basis for live streaming TV over the internet (OTT). Unrelated, but important for me, TV stations in my country (Serbia), forbid rebroadcasting but they are available OTA, with a DVB-T2 receiver and an aerial/antenna. However, some cable providers cannot show the programs, and if they do show them, they have negotiated special licenses for that. How do I make a case that by being OTA Free, and not DRM'd, I, by not changing the contents of the broadcast, should be allowed to rebroadcast to people for no financial gain to myself or the TV stations (other than the ads the TV stations put inside of the broadcast content) (all to serve as an experiment)?
Note, you still have to deal with TV ads and there is no scrubbing. I wonder if you could make your own client to this and get DVR functionality? Actually since they are just grabbing over-the-air signals, I wonder if there is a SDR project for this.
More likely they'll never expand beyond 500 "experimental" accounts and shut it down once their research is finished. I don't intend this as a criticism of the researchers; the more academic and experimental you make something look the less likely you will be threatened/sued into compliance.
We're working on a paper for submission to a research conference soon, so every viewer gives us helpful data for the system to learn in-situ and improve its performance over different kinds of Internet connections and congestion situations. (We are randomizing connections to get either our in-development algorithm or a bunch of competing ABR and congestion-control schemes so we can properly evaluate, so if you get a stall or bad quality...... we'll blame that.)
All the source code is at https://github.com/StanfordSNR/puffer and is or will be released with an open-source license. The study is led by my doctoral student Francis Yan, along with Sadjad Fouladi, Hudson Ayers, Chenzhi Zhu, and my colleague Philip Levis. The goal is to train video-streaming algorithms (bitrate selection and congestion control) online, continually, in situ, to minimize stalls and deliver the best picture quality across a diverse range of real-world Internet connections and users. This is from some of the same people who brought you Salsify (functional video compression for lower-latency real-time video: https://snr.stanford.edu/salsify), Lepton (distributed/parallel JPEG compression: https://github.com/dropbox/lepton), Mosh (functional terminal emulation for mobility: https://mosh.org), Remy (machine learning for congestion control; http://mit.edu/remy), and the Pantheon of Congestion Control (https://pantheon.stanford.edu).
There are some unusual details of the Puffer system to make experimentation easier (and, we hope, performance better): the ABR algorithm is server-side, and video is streamed continuously over a WebSocket and the client gives asynchronous feedback, instead of using DASH HTTP request/replies. The tcp_info struct (including the delivery rate estimate) from the congestion-control layer is plumbed through to the ABR, and to the neural network that tries to predict "how long will it take to send a video chunk of length y," so it's weakly cross-layer. The "transmission-time predictor" is probabilistic instead of simply producing a point estimate. We encode with libx264 and then measure the SSIM of every encoded chunk to calculate a reward for each possible stream (instead of making the assumption that more bitrate is better -- the correlation is not so good when you consider different times in the same video). And we encode 10 variants of each channel, so more than typical. And everything is done at 60 fps. Of course some of these details mean these algorithms are not going to be deployed any time soon on a production CDN or streaming service, but the hope is to demonstrate the value of some of the ideas in an academic setting for future commercial deployment. The use of WebSockets, MSE, and Opus is why this doesn't work on Safari or on iOS.
FAQ here: https://puffer.stanford.edu/faq/
Also I'm happy we were able to make some of the Grafana monitoring public: https://puffer.stanford.edu/monitoring