Hello, and welcome back. This week we are going to be talking
about image and video inpainting. In contrast with the topics that we
described last week, where we performed a description of fundamental tools to solve
a number of problems, like an isotropic diffusion, active contours, and even
image inpainting. This week is about a particular problem.
It's not about tools for multiple problems,
it's about solving one particular problem in image and video processing.
Of course, some of the tools that we are going to be using for this might be used
for other techniques, and we often learn from the solutions of
a problem towards another problem. But it's a very important difference
between providing for the mental tools for solving multiple problems, and
solving one particular problem. So what is image and video inpainting?
And let's start with image inpainting. Image inpainting is basically the art of
changing an image in a non-detectable form.
So here we see a painting that has a lot of degradation, and here is the
restoration of that painting. This was of course done by professionals
manually. It's a painting, it's not a picture, was
not done in the computer, although the computer can help as well, as we're
going to see it later on. When you go and see this picture in the
museum, you'll believe it's the original picture.
Some experts actually might be able to notice the difference, and be able to see
the regions that were actually inpainted or filled in with information from
different regions. But the basic idea is that in inpainting,
we show you an image or a video, and we make you believe that's the original one.
In this case, it's to solve degradations. Inpainting, it's sometimes also called
fill in, because we're filling in the regions of
problems, the regions of deterioration with
informations from this surrounding. And we are going to see that, sometimes,
with fill in with information from the near areas sometimes with information
from far away areas from the whole image. There's even techniques that do image in
painting using database of images where you look for similar things.
So, this is one example of what image in painting needs to be done.
Remember, it's modifying an image in a non-detectable form.
And when we say non-detectable, we say at least by the non-experts.
Here is another example. A lot has been done to this image, basically to
transform it into this one. Of course it has been inpainted, you see
these regions have been filled in. But also there were color corrections and
other things, but this illustrates yet another example of image inpainting.
Look at all these regions that have missing information and that were
basically restored, inpainted, filled in with nice information that looks like
natural in the surrounding region of the image.
Of course, it's natural to extend these to regular photography, in particular
when we were talking about pictures, and not digital pictures, but analog
pictures, and here we see examples. Some of examples like here have a few
additional restoration techniques, but we see these.
For example, the picture was turned into two pieces, then was put together and to
repair that, you basically see that it was inpainted.
A region of missing information or undesirable information was filled in
with other types of information. We see here the same.
There is a scratch. Now the scratch is gone.
And here there is a similar example like this.
Basically, the picture was folded, and when you unfold it, you see this kind of
scratch in the middle and then here its gone.
So, these are all examples of image impending to restore basically the image
to a much nicer looking image. Of course, this is an example that we saw
in the first week, where image inpainting is done with a completely different
target. In this case, and this illustrates the
idea of modifying an image in an undetectable form.
In this case, we see nothing, nothing wrong with this image, that we know, if
you remember from the first week of class, that basically this was the
original image. So what was inpainted now is the object
that the user wants to remove from the image.
So basically, this region becomes the region of missing information that we are
going to inpaint in, and we need to recover the water, and we
need to recover the column. So we need to recover texture,
we need to recover structure, straight lines and water that is
basically non-straight lines. So very different types of informations
need to be filled in, inpainted, in order to solve this problem of removing
objects. And we discuss in the first week, this is
used all the times in the movie industry, I show you a few examples of objects that
need to be removed in the movie industry. Here is yet another example from this
era, where basically we see here, a picture of Lenin, and a picture, and next
to him there is Trotsky. And here, basically, Trotsky has been
removed from the picture. So once again an object has been removed
from the picture. And, yeah, you can have multiple examples
of these. For example in this website, or the website that was linked in the
previous slide. So again inpainting in order to remove
objects that from the image or from the video.
I want to explain something that inpainting, in its basic form, cannot do.
I don't want to, you have the wrong understanding that inpainting is an
extremely smart technique that can fill in with whatever you want.
So let me use this example to illustrate that.
Here, this person was removed, and then it was inpainted.
So when you look at this image, it holds everything that inpainting is about,
modifying the image in an undetectable form.
But look what happened here. When the person was removed, two chairs
were brought in. This is done by an expert.
It's not done by the computer. There's no way for the computer to
understand that what you want here are chairs.
Maybe in a very, very advanced technique, when you basically provide a lot of side
information, you could, in the computer, with programs
like Photoshop, cut chairs from here and put them here.
And this is the kind of stuff that was done here.
So it's a cut and paste, and we are going to see how to do that.
But basically, inpainting, does not have that high level knowledge of what needs
to be there. The computer and the user can interact to
basically bring that into the scene, as we're going to see later on.
But if you don't interact, if you just say, do something with this
region, the computer does not know what is that
you want, why is that you want chairs there,
and not, for example, green, not, for example, some of the background.
But with interaction with the user, we might be able to do this,
kind of a cut and paste. So, this is about image inpainting.
In the next video, I'm going to show you that we are actually familiar, in nature,
with image and video inpainting, and that's actually very interesting.
See you in the next video. Thank you very much.