Hello everyone and welcome back to the computer vision lectures.

This is lecture 2 and in this lecture we are going to talk about how images are formed

and a little bit about light.

So let's jump into it.

Before we go ahead, I want to ask you an important question.

Think about why is vision difficult?

Why is it difficult to understand an image?

Let's say you are in a living room and you take a photograph of your living room.

You will see in the photograph all the different objects that are lying around in your living

room, different people sitting there, what color those objects are.

You will be able to see every aspect, the light, the geometric shapes, the relationships

between the objects and it's very easy to understand for humans.

So why is it so difficult?

It is important to know that computer vision usually works in inverse problem domain.

What do I mean by inverse problem domain?

In inverse problems, you are given an image or a signal from real world and you have to

infer different characteristics of that object from the real world.

For example, geometrical properties like 2D, shape, size, things like that.

And also you have to infer appearances, 3D information from images and so on and so forth.

The problem is that there is not enough information here.

Why do we say that we don't have enough information?

Because we just have only a single image.

And from that we have to infer all these characteristics of the real world.

So it becomes a bit difficult in terms of crunching numbers.

Another problem is that we don't have most of the problems in computer vision are ill

pose problems.

What I mean by ill pose problems is that there are insufficient examples to match the unknowns

that are present in the problem.

Let's look at a previous example of the real world image.

This is a known entity that we have.

From this we have to infer a lot of different aspects as we saw in the previous slide.

And these are the unknowns and there are a lot more unknowns.

This is a very rough approximation of this ill positeness of computer vision problems.

Does not necessarily mean that it is going to be like this every time.

So before you jump into any problem of computer vision, you have to really study what aspects

of the problem are available, what prior knowledge you already have, what assumptions you can

make, what constraints have to be added like physical geometrical constraints.

And then you can go ahead and then you will know or have a better understanding of what

exact unknowns you need to be inferring.

Okay, so let's move on to ask ourselves what is an image really?

For us an image could be random values between 0 to 255 and a matrix of those values and

you see that those values in an array or in an image form and then that's an image for

you.

But does it really that easy?

Is it that simple?

How many values can it take?

How many values can one pixel take?

For example, when you generate a random image of random values, it can look something like

this.

But do you really consider this as an image?