I assume something similar to sponsor block, some algorithm to identify ad segments and some user feedback to confirm.
Unless I’m mistaken as to how sponsor block works?
People will watch the videos, report the segments that are sponser slots, and then when people watch the video they can upvote or downvote the accuracy of the report.
In stream ads would be a hard one to tackle because YouTube would likely inject them randomly into the stream to boost engagement (readas, prevent people skipping them easily).
if they were randomly placed, then couldnt you have a sponsor-block type system where instead of the ad segments being marked and skipped, information about the video is externally stored somewhere (like perhaps a really low res screenshot of the video every couple seconds, or some number generated algorithmically by a frame of video), and the results should be the same for all users for the actual video part, but if the ads are placed randomly, the ad section will suddenly not match the data other users had, prompting the video to skip until it matches again (with a buffer included if they remove the ability to move forward)
Take two copies of the same video, diff them and only keep the parts that match.
We can also build up a database of as signatures to automatically identify them without requiring a watermark - we already have the technology to do this for detecting intro sequences for skipping.
In that case the ads are video only, no clicking on them, including to skip or anything else. So it would be detecting that trying to change where you are in the video doesn’t change anything (and exclusively playing via your 3 second buffer)
I assume something similar to sponsor block, some algorithm to identify ad segments and some user feedback to confirm. Unless I’m mistaken as to how sponsor block works?
Sponser block works via user input
People will watch the videos, report the segments that are sponser slots, and then when people watch the video they can upvote or downvote the accuracy of the report.
In stream ads would be a hard one to tackle because YouTube would likely inject them randomly into the stream to boost engagement (readas, prevent people skipping them easily).
if they were randomly placed, then couldnt you have a sponsor-block type system where instead of the ad segments being marked and skipped, information about the video is externally stored somewhere (like perhaps a really low res screenshot of the video every couple seconds, or some number generated algorithmically by a frame of video), and the results should be the same for all users for the actual video part, but if the ads are placed randomly, the ad section will suddenly not match the data other users had, prompting the video to skip until it matches again (with a buffer included if they remove the ability to move forward)
You don’t need anything so complicated.
Take two copies of the same video, diff them and only keep the parts that match.
We can also build up a database of as signatures to automatically identify them without requiring a watermark - we already have the technology to do this for detecting intro sequences for skipping.
In that case the ads are video only, no clicking on them, including to skip or anything else. So it would be detecting that trying to change where you are in the video doesn’t change anything (and exclusively playing via your 3 second buffer)