Video/audio/image "enrichment"
Booted Cat
"pad" a poorly rendered video with details suggested by statistical
methods. That is to say, if we have a large video corpus and we can
find other high-resolution videos that have similar "contexts" with
our poor video, then probably we can borrow the details from such
videos to pad the poor one.
This is already done for texts: there's a program called "WhiteSmoke"
that suggests modifiers to your text. Microsoft Word 2007 also has an
"English Assistant" that suggests phrases that contains a word you
give (e.g. "search" -> "diligent search" / "thorough search" / etc.).
Regards,
Yao Ziyuan
Torben Ægidius Mogensen
> Seen a low resolution video just now. Leads me to wonder if we can
> "pad" a poorly rendered video with details suggested by statistical
> methods. That is to say, if we have a large video corpus and we can
> find other high-resolution videos that have similar "contexts" with
> our poor video, then probably we can borrow the details from such
> videos to pad the poor one.
Yes. Similar methods have been used to repair "holes" in images (such
as scratches), often using statistical data from other parts of the
same picture rather than a corpus of similar pictures.
But you can only do so much. If you, say, have a lowres picture of a
face, you can generate a hires picture, but it may not look at all
like the "real" face. The Hollywood cliche where experts are given an
extremely blurry picture of a face and through spending days (to keep
up the suspense) of CPU time can produce a crips and easily
recognisable face is just that, a cliche. You can't recover lost
information, you can only add information that _might_ me similar to
what was lost.
With movies, you have the additional advantage of being able to use
the time dimension: Assuming a fixed camera, large parts of the
picture won't change between frames, so you can take information from
neighbouring frames. Even if the camera (or background) is moving,
you can often identify regions that move by a specific delta and
inside these regions you can use date from neighbouring frames at this
delta. Such methods are often used in movie compression, where you
essentially try to "lose" information that you know that you can
recover again.
Torben