Original subtitles

1

Hello and welcome back to the course on computer vision.

2

And today we're going to talk about cascading.

3

So previously we discussed about the show on classifiers and how they're constructed from Week classifiers

4

and put them together.

5

And now we're going to talk about another hack that the burial of Jones algorithm applies in order to

6

speed up the process.

7

And the hack is called Cascadian schematically it looks like this.

8

We take a window that we're evaluating on our.

9

So that's a window that is sliding across our lives and we look for the first feature in all or list

10

of features.

11

So for the most important feature and then we see if it's present on the image or in that same window

12

if it's present in some window or not if it's not present in the same window then we reject that window

13

completely and we don't even evaluate the remainder of the picture.

14

So if for instance this is the most water feature and we find out that it's not present in some window

15

then we're not even going to look at these.

16

Then if you say the way to think about it is if if there is no nos an image then it's just not going

17

to be a face like a frontal face has to have a nose therefore what's the point of looking for eyes and

18

cheeks and eyebrows and so on if we're you know that there's no nodes.

19

Then if this feature is in the presence of if there is a nose then it is time to walk for the second

20

feature.

21

Now we're going to evaluate if the second features in that sampling them if it's not then we reject

22

the subway for instance if there's a nose but then whatever this is there's this looks thing an eyebrow

23

but it's upside down so maybe there's no Water-Lily.

24

Whatever this feature represents whatever the algorithm was looking for there.

25

It's not present then we're going to now reject the reject image because it has a nose and doesn't have

26

a lip so on your face.

27

So then if certain value the third feature if it is present if that should also happen then and without

28

talked.

29

And so on.

30

So if servies is not present then we reject and move on.

31

So that's the simplistic diagrammatic kind of explanation.

32

In reality it's a bit more complex.

33

In reality of course it's very risky to rely on just one feature because maybe it might not be.

34

Maybe the shade is in such a way or the lighting is different or maybe just so for some reason wasn't

35

detected as can be lots of different circumstance so what it actually does is in this first step it

36

doesn't look at the one feature it looks like the top five features with the top 12 inches.

37

Well let's say let's say Tokai for argument's sake it looks at top five features and then decides based

38

on them so if if they're not present if for instance none of them are present then or some portion of

39

them are present like the details the details are not important right now what is important is the whole

40

concept that regardless if you find identifies based on the first five features it takes one of two

41

or three or four or five and it's based on them that this window this some Windu is pointless had to

42

keep looking at it's not a face it won't continue then here it will look at the next 12 here or look

43

at the next 25 or something.

44

So every time because the features are becoming less and less important less and less prominent is taking

45

more and more of these features.

46

But nevertheless is basing its decision on whether or not to continue going forward with this on the

47

at the at that point in time all the features it's looking at.

48

And if it is not happy with one and seeing things that it might now like if it decides that over here

49

that right now this should be the these features and all this many of these features that are not there.

50

This is not a face.

51

I'm not even going to evaluate the rest.

52

And that is called cascading.

53

It helps really speed up the process.

54

So in visual terms let's say that these are all features that we identify.

55

Again we'll go back to the simple example of just getting them one by one rather than in batches.

56

And so what we're going to look at is we're going to look at the subgroup over here.

57

And so let's say we look at this feature and we can see that it is present.

58

That's the eyebrow over here.

59

Dark and bright.

60

So this will pass so that that passes.

61

OK so it goes from here to here.

62

Good.

63

Second Future looks for this.

64

You can see that in the eye.

65

OK it passes then goes to this and maybe in this case it actually can see that.

66

So this is technically this feature is reserved or was originally planned for the lips so light black

67

lights a light dark light and it would that would have identify liberties we've got an intuition that

68

actually eyeball you can see the same thing you can see light dark and light so it can mistakenly actually

69

accept the this feature is here and would say oh this features here so we do have the lips and really

70

we don't.

71

But OK so even if that passes it gets to this future and it cannot identify this feature because this

72

feature for instance like nowhere on this image you can see like light and dark until dark and you can

73

see the light and dark of light but not the other way around the white and black like that.

74

So it was actually for this part of the nose as far as I remember.

75

So at this point it would at this point it would reject I would say this feature features are present

76

so featuring them for the present.

77

They reject the sub window completely and it wouldn't even evaluate this feature.

78

And as you can imagine you can get kind of 100 features or how many features after that it would just

79

not evaluate them any of them.

80

And even though evaluation takes like split seconds for each one but when you add them up because there's

81

so many of them it takes time and plus you have to go through this image lots of time with some windows.

82

And so it saves not just time devaluing the Swan But lots of features all the features that come afterwards.

83

And it just rejects this whole sub window and moves on.

84

And that really speeds up the process and that's called cascading.

85

So on that note we're going to wrap up today's tutorial.

86

I hope you enjoyed it and I look forward to seeing you next time.

87

And until then enjoy computer vision.