Okay, hi everyone to another session of our Computer Vision Zoom.

Today there is a demo talk given by Vanessa Wirth.

She is a PhD student at the Department of Visual Computing, sorry Department of Computer

Science Chair of Visual Computing and today's session will be focused on traditional methods

in computer vision from graphics perspective. She will show some demos that are applications

which have applications in the graphics. So I welcome Vanessa and thanks to you again

for doing this demo and the stage is yours.

Yeah, thank you. Hello everybody. As you might have already heard, I'm from the Chair of

Visual Computing and since computer vision is a pretty interesting and training field

nowadays you will see that multiple academic chairs nowadays will get in touch with computer

vision and so am I. So a little bit about me beforehand maybe I'm focusing primarily

on 3D reconstruction. So I started with 3D reconstruction of static objects, for example,

furniture that doesn't really move and nowadays I'm focusing on 3D reconstruction of non-rigidly

moving objects such as us humans, for example, and I'm trying to achieve faster movement

reconstruction basically. So that's what I'm doing nowadays and today you will also hear

a lot about reconstruction of course because that's my speciality and today we are focusing

on traditional methods. So let's start. You already heard something about passive image

acquisition as I saw on the slides. So I'm presenting you today the Metashape program

which is basically a collection of passive algorithms. So just a short recap, you might

heard about the three stages that you need to find correspondences. If you want to reconstruct

something you need to build up a feature descriptor and then you need to sort of match the descriptors

together by nearest neighbor matching or something. So you're matching correspondences basically

and that is also what Metashape does but in a really professional way. I actually did

some digging about Metashape because it's a proprietary program. You don't really actually

know what they are doing but I was searching on the forum what algorithms they use. So

here's a short description of what I know so far. Metashape is a photogrammetry software.

It is primarily used by archaeologists because they of course want to reconstruct things

they see really ancient stuff and so they use this quite often but you can also see

it in the movie or in the game industry. So for example the Unreal Engine uses it, Halo

4 uses it. You can also see it in Mad Max actually because the good thing about this

reconstruction software is you can take a photo of the scene if you find it beautiful

around the world and then you can use it in the movie industry for example. So that is

actually what Mad Max did. They did a reconstruction of the desert as you can see here and those

objects here are all reconstructed. So you can see the scene is reconstructed, the vehicles

are reconstructed and then they just did some blending and lots of graphic stuff to make

it look really really beautiful as you can see in the picture on the bottom. So that's

actually what also happens in the movie industry. And as I already said I did some digging what

they might do in their algorithms. It is a collection of structured promotional algorithms.

You already heard about that and it definitely uses SIFT descriptors. They also say that

to estimate the parameters of the camera in the current image you are seeing when you're

taking a photo they use something that is called bundle adjustment. You might not heard

about that yet but you heard about the single stages that are actually involved in bundle

adjustment because you heard about the fact that you sort of need to find points in the

3D space and you also heard about the fact that you need camera parameters to do the

matching from the 3D points to the 2D pixels. And that is actually what bundle adjustment

does. It does everything as a whole thing. So that's also why it's called bundle adjustment

because it does everything in a bundle. It estimates camera parameters and also the 3D

coordinates of the points that you can see in an image as a whole in a collection. So

that is the algorithm they use. And I also saw that they to get the final 3D surface

they use multi-view stereo because of course they're also taking multiple pictures. You