Meeting minutes
Agenda
cpn: Main topic will be impacts of AI on MEIG's mission.
… Also want to raise TPAC and potential meetings.
… Anything else anyone would like to cover, or announcements to make?
Nigel: We published a new version of IMSC recommendation a few weeks ago
Nigel: If it needs announcing, TTWG published IMSC Text Profile 1.3 two weeks ago
TPAC preparation
cpn: [shares screen]
… Last year we had a morning dedicated to MEIG topics, on the Monday.
… In the afternoon we had a joint meeting with TTWG,
… and later in the week a joint meeting with TTWG and APA WG
… At the moment, we've been asked to confirm how much meeting time we want.
… Then we can figure out more details of the agenda and so on.
… My suggestion is that we do the same as last year,
… unless anyone thinks we need maybe additional time for specific topics.
… If the TTWG wants to we would be happy to have a joint meeting.
… The APA WG joint meeting has generally been quite productive as well.
… I propose we request the same structure as last year.
… Thoughts on that proposal? Or other suggestions?
Nigel: Sounds good to me
… I'll check with TTWG on Thursday this week
cpn: I will assume we are comfortable with that proposal and request those timeslots
… I've also created a GitHub issue where we can collect items for the agenda.
w3c/
cpn: I welcome requests for agenda time or topic suggestions in that issue.
Nigel: AI is a hot topic, and there are likely to be concerns or questions or proposals for changing how people consume media on the web, browsers with local agents
… It might be prudent to allocate time for that as a specific topic
cpn: I welcome that, thank you.
AI impacts on MEIG mission
<Roy_Ruoxi> https://
cpn: I noticed Roy you have been editing the Web AI roadmap. Can you talk to us about it?
<Roy_Ruoxi> https://
Roy_Ruoxi: Above is the repo where we're working on it.
… And the document where we're talking about W3C technologies.
… First we are describing existing W3C technologies to support AI, things like
… WebGPU, WebAssembly and WebRTC, and some data things like RDF, JSON-LD,
… that AI agents might use to understand web content.
… There's also CG work like AI Agent protocols.
… WebMCP is worked on in the Web Machine Learning CG.
… There are some security and privacy and ethical considerations.
… This year we realised that not only our technology supports AI in the web platform
… but also we should consider how AI effects our existing technologies.
<Roy_Ruoxi> https://
Roy_Ruoxi: All the issues were set up by Dom with me working with him.
… We asked each WG and IG if AI has impacts on existing standards or if anything
… needs to be adapted due to emerging AI things.
… For the MEIG we would like participants to think if there is any impact to our existing
… work, or if there are gaps between web AI technology and our existing work.
… That's the purpose and background, why there are a lot of issues in the webai-roadmap repo.
cpn: Thank you. The issue asks about the impact of AI technologies on the mission of the group,
… i.e. the scope.
… We have a mission statement in the Charter
… It looks like the mission is the same, but AI brings a new topic of conversation and focus
… to consider, but from a charter perspective it's within the current remit.
… We can go into some specifics soon.
… There are wider implications in terms of the impact of AI media creation on the web
… as a whole, which is a slightly different topic, but interesting.
… Media is being increasingly created with AI, and that affects what people are able to create,
… with beneficial and perhaps some detrimental effects that might be interesting
… to analyse also.
… It has led to media requirements about content labelling, from regulators for example,
… which might drive the need for technical standards for labelling.
… For the feedback you're looking for, do you want specific examples or use cases?
… What input would be helpful?
Roy_Ruoxi: I think everything is good!
… I'm not an expert on all the aspects. With regard to M&E I guess it might have an
… impact on a lot of things. The use cases would be helpful, so we can integrate them
… and see if there's anything W3C needs to do.
cpn: In practical terms, is it better for us to collate that input here and send it to you,
… or comment directly in the GitHub issue that Dom created?
Roy_Ruoxi: You can comment on the issue directly, or if you don't want it public then
… contact me privately.
cpn: OK, sounds like that might lead to potential future work.
Rob: Thinking about content labelling and creation, AI is basically pattern
… recognition and media is a big pattern of audio and imagery, which seems like a rich
… field to explore, with many issues to explore, looking at it from the other perspective.
cpn: Good thought.
Rob: And the impact on security, privacy and copyright.
kaz: So-called "AI" today is based on generative AI
… we should be clear about which use case and which parts and which industry
… is related to W3C standardisation.
… If the MEIG would try to handle updated recent version of metadata handling using
… so-called AI services that is fine. Any kind of use cases could be applied.
… For example, media metadata and format discussions e.g. proposed by Dolby
… should be our first priority from the MEIG viewpoint.
cpn: That's consistent with the scope of the Charter.
cpn: Thank you
Nigel: This came up in TTWG, we looked similarly at our mission statement. From the point of view of timed text data formats, there isn't really any change to make
… either based on creation or consumption of timed text data if AI is used
… But we also noticed that there are developments that could impact the content of specifications we write. Some organisations add data to subtitles and captions, e.g., tone or emotion or loudness of sounds
… When presenting timed text on screen, with user customisation, allowing those dimensions to modify how text is presented: text size, font, animation
… to try to convey those dimensions and give a richer experience to audiences. We thought this is a good moment to open things up to allow new standardisation requirements to be supported
… Some companies do this by hand, takes 1 hour for 1 minute of content. That's not sustainable. They created an AI product to do it more quickly
… There are use cases AI might enable practically that we haven't captured. In TTWG we could try to gather those requirements. We want to ask MEIG, in a place with less IPR commitments, set up a TF to investigate and see if there are requirements to bring back to TTWG
… Invite companies to come talk to us
cpn: Any reactions from anyone?
… I think that would be quite welcome, and if we can attract contributors to help us
… do that then it would be more likely to be successful.
… Personally I'm supportive towards that.
… Logistically it is easy to create task forces.
… We just need a resolution in the meeting.
… Practically we need a task force chair/moderator/facilitator.
… Do you want to try to do this today or take it up offline?
<RobSmith> Scope
Nigel: Suggest drafting something to cover scope, timescales. I propose to write something and come back with something more concrete
… Please get in touch if you have ideas
AVS use cases
Song: Mr Zheng is from PCL, joining W3C. He presented to Chris and W3C staff at the AC meeting in Hangzhou
AVS Introduction
Jianhua: AVS has 100 members, it makes specifications for audio and video technologies
… AVS3 video standard recently published
… AVS is now working on AVS4, based on AI
… AVS has developed over 20 years, published in China and as IEEE standards
… In 2022, referenced by DVB
… DVB specifications reference AVS3, media and smart media transport
… Three years ago, combine AVS technology into DVB-I framework. AVS3 P2 video, +P6 media format. AVS multiview, AI stereoscopic video
… In 2022 we did a pilot at the Paris Olympics, using SMT in the DVB framework, delivering content across screens: TV, STB, phone, tablet, XR glasses
… One signal delivery, received across different terminal types
… Launched verification testing for terminal presentations, and delivery on different frameworks synchronised on the terminals
… We use SMT enabled interactive broadcasting. For multiview we have multiple camera interactions, select the camera angle to receive
… This year in the Milano olympics we did another live streaming verification using this platform. This time we had 2 live channels and 12 non-live channels
… We set up 7 types of terminal as receiver
… China Media Group in Shenzhen
… Verification for live and non-live streaming, and combined with content protection. Also AI enabled live caption translation
… In IBC 2024 we set up a demo, and last year at DVB World we showed use with wearable devices
… Free-view technology with 24 cameras at the live field, data compressed equivalent to the 4K signal
… Users can change the view angle on the TV using their phone or remote control
… Demo of Milan winter olympics. Three types of terminal, using the same UI. 4 channels where user can select
… We want to build future TV experiences. Presentation across TV and other devices. Interactive appearance so user can change the viewpoint
… We have AI technology to bring this interactive broadcasting
Use case study: Volumetric video
Song: Additional use cases for TV and mobile
… Emerging technology like volumetric video and MPEG MCVC, Using this for commercial applications in China
… [Shows demo]
… Mirroring between the TV and the small screen, controlled by hand gesture
Use case: 2D and 3D conversion
Song: It's another 3D user experience. Uses low-end 3D glasses, but the processing and compute is in the set-top box and in the cloud. We can use native application or HbbTV, or MSE in the set top box even though there's still a performance gap. We're working on that
… Uses VVC and DASH, and AVS and SMT for transport
Observations
Song: Volumetric video frame interpolation. Denoising using AI. AI agents (ChatGPT, Claude, Doubao, Qwen, etc) evolving from text based to A/V multimodal
… There could be higher demands on web platform AI capabilities
… Neural networks replacing traditional codecs. WebCodecs has support for H.264/AV1/VP9 with browser encoder and decoder interfaces. Using the NN model codecs isn't ready yet
… WebGPU supports parallel data processing
… For volumetric video, recoridng information of object information in 3D space
… Data traffic will be very high, up to 720 Mbps for 30 fps, so 20-30x
… For real time interaction, the user wants gesture interaction with instant rendering, which is a challenge for set top boxes using HbbTV and MSE
… 3D point clouds, we use semantic processing with the media. Neural network rendering using WebNN and NeuVV for high fidelity rendering
cpn: Wow, thank you, that's amazing!
Chris: Thank you for the presentation and thorough analysis
Rob: Regarding data rates, are you proposing this is all delivered then composed in the set top box? Broadcast all 24 channels, or does the user select which one and only stream that selection?
Song: Broadcasting all the signals would be expensive. The normal view angle is like traditional broadcast, then when there's a highlight, you can use the remote control to select. The baseline streaming is a typical HDR video. The server retrieves the different angle stream from the cloud.
… In the reference example, they use 24 cameras to distribute the Games. Switching between angles may not be as smooth as expected, so use 36 streams or for important games with premium subscribers, use 72 cameras.
… MSE has some performance limitations. So we recommend a native solution
Jianhua: The free view use cases uses 34 cameras in the live field, using a different technology.
<Zakim> nigel, you wanted to talk about TT
Chris: Any WebCodecs requirements to bring to Media WG, or similar for WebGPU?
Song: The problem is that set top boxes are middle or low end devices, but the performance doesn't meet the requirement. I suggest we discuss the performance requirements
Chris: I remember the Cloud Edge Compute work in the Web & Networks IG
Song: We finished the design for that, but it didn't get adoption. The implementation chain is too long, hard to convince the broadcaster, CDN provider, and source of program to adopt
… So if we're going to implement the technology we need to make the chain as short as possible
Chris: Next steps? Happy to use more meeting time.
Song: I plan to provide use cases for the AI media roadmap for W3C
Chris: I'll create a GitHub issue to collect use cases
Rob: Any proposal to tag content, add timed metadata, to identify the person on screen. e.g, athlete at sports events
<RobSmith> next meeting 7 July