Hello everyone and welcome back to Compatibhizion lecture series.

This is lecture 11 part 1.

There will be another part to this lecture but this is the final lecture for our course

for this semester.

So let's just jump back in and start from where I left.

Until now what we have discussed in the last lecture specifically we discussed about a

structured light which was using the same geometrical construction as optical flow finding

dense motion estimation.

And we used the same ideas same geometry and applied it using structured light algorithm

where we can use or where we can estimate depth or the shapes of the objects under consideration

even when the surfaces of objects are quite uniform.

So that was the main advantage of using structured light approach.

In one of those instances we discussed also about portable laser scanner.

So in this scanner what does the scanner generates is like a point cloud.

It generates in 3D space different points that are it localizes specific points in 3D

plane where the which represents the distances of the points on the surface of the objects

being scanned.

So in this case when after scanning the face of this person when the 3D point was generated

the 3D point cloud.

So this is a collection of all these 3D points in real life and therefore it is a point cloud.

We can consider it as a point cloud or a 3D depth map or 3D shape of this person.

Here you can see that there are a lot of missed points and these points which are also estimated

with high accuracy however are not smooth.

So they are not smooth in the sense that we can't have a very high resolution.

We can have a very high accuracy for the localization but not high resolution because there is a

limit of limit to the number of points that we can use as we saw in the case of using

connector.

So I want to build up on the same idea and move forward and talk about surface reconstruction.

In this another task or another topic of computer vision here what we do is using these different

3D scans or 3D point clouds of the same object from different angles, views and stuff like

that from angles, views and positions and exposures and all those different parameters.

We generate a holistic more dense very accurate surface reconstruction.

So it's not always possible to get these 3D point clouds with one imaging method.

So instead of that what we do is we generate multiple such scans and then we merge those

scans.

So given these dense set of 3D points from multiple scans or stream of left images or

reconstructed from multi-view stereo our main goal or challenge or the final product or

result of surface reconstruction is to reconstruct a closed surface of the object.

What do we mean?

So we have subscans or multiple scans like this the green and the red ones and then we

merge it to form a more continuous more solid more rigid surface of this object.

In this case it is a rapid.

Surface reconstruction has multiple problems in it.

The first problem to address is that there needs to be multiple scans and they need to

be aligned.

So in order to generate a more holistic more rigid surface from multiple scans we need

to align them.

So that's the first problem to address.

The second problem to address is that after merging all those points in proper alignment

how do we convert this dense point cloud to a surface here on the right hand side.