Open dobkeratops opened 5 years ago
Is the video still available?
I always get "video not available" when trying to access the video...but not sure whether that's some sort of country restriction?
i copied the link wrong (sorry still got this awkward situation with logins/half broken laptop).. here it is again
https://www.youtube.com/watch?v=FlgLxSLsYWQ&t=22m10s
the part at 22:11 is the part I wanted to draw attention to - art via images <-> vocabulary .. why I think we need 1000s of labels for rich descriptions, to open these imminent use cases.. but the whole video is worth watching too.
Whatever machine-learning techniques we have, until computers cross that 10petaflop range (whatever it is) human markup will always be economical and needed
ithink ifound the paper they might have been talking about there,or something closely related here's a lecture:- [SAGAN] Self-Attention GenerativeAdversarial Networks | TDLS https://www.youtube.com/watch?v=FdeHIC4QiqA
I'm going through finding loads of reerences at the moment (space between human capture, machine-learning, applicatino to CGI)
the part at 22:11 is the part I wanted to draw attention to - art via images <-> vocabulary .. why I think we need 1000s of labels for rich descriptions, to open these imminent use cases.. but the whole video is worth watching too.
wow, that's really amazing, thanks for sharing!
I totally agree with the presenters outline of where we will be in 5 years from now. There will definitely be more AI powered systems in our everyday lives. But I hope that this technology is available for everybody then (hopefully we will see more open source projects that are built around publicly available data) and is not limited to the big companies...
why I think we need 1000s of labels for rich descriptions,
totally agreed!
I really hope that the properties system and the soon coming scene labels will increase the value of the dataset further. :)
btw: I am in the middle of migrating my workstation from Windows to Ubuntu (again). I was running out of disk space (150GB was way too low in the first place), so I had to replace my harddisk. Thought that was a good opportunity to switch the OS as well. I think I should be done with the migration in the next 2-3 days...then I'll start with adding a bunch of scene labels. (I'll use the list you posted here a while ago as starting point :))
video touches on areas that interest, https://www.youtube.com/watch?v=FlgLxSLsYWQ
22:11 onward "Generatedimages from text"
this specifically touches on why I'm pushing for rich vocabulary.. describe each objet and its pastas accurately as possible - I think graphical use cases will take thousands of labels at least.Hopefully the properties system will help .. we can think if we can generate aliases or whatever
They touch on how nvidia,(adobe?) etc have NNs for this.. an open source dataset could be there for the blender based CGI communty