Original subtitles

1

Hello and welcome to the practical applications of module to object detection I'm super excited to start

2

this module for two reasons.

3

First one is we are taking things at the next level now.

4

As I told you open TV is not the most powerful model and the model we will implement in this module

5

is much more powerful because it is based on deep learning and neural networks that computer vision

6

where the computer will have a brain.

7

That's exactly what it means.

8

And the second reason is that we have an exciting challenge.

9

I will show you a video of a very cute dog bouncing on the field.

10

And our challenge will be to detect the dog will be to implement some program that will detect the dog

11

in the video.

12

So it's good that you see several ways of doing some computer vision in the first module you learn how

13

to do some face detection through a webcam.

14

And now you're going to learn how to do some object detection on a video directly.

15

Now before we start I would like to say a big thank you to this developer here next to Groote.

16

That's a picture of him in a horseshoe.

17

He's the creator of the PI torch implementation of single shot multi-book detector that we're going

18

to use in this module.

19

So thank you very much for sharing this and make it open source.

20

We actually tried several object detection models we tried the first are CNN the yellow open CD and

21

the SSD and we obtain the best result with the single shot multi-book detection.

22

Not only we obtain the best result with this moral but also if you look at the paper you will see that

23

on the tested cases the single shot multiplexed detection model beats yolo and fester are CNN.

24

So that's why our choice for the ultimate objective texture model of muchall 2 was single shot multi

25

box detection and the best implementation we found was from this developer Max agreed.

26

So thank you so much.

27

Thank you for sharing this pre-trained moral it's actually a pre-trained moral that was trained to detect

28

between 30 and 40 objects including cars dogs horses ships boats planes and more.

29

So a very useful not all that you could use for your own business problems.

30

We're going to go inside the SSD in this module and we're going to learn how to use it and how to detect

31

any object on any video.

32

So that's going to be a pretty exciting module.

33

I can't wait to show you this video of this Doug.

34

I really like this Doug it's actually Carol who filmed this little dog with a drone.

35

So the first thing we're going to do now is we're going to open Anaconda because I want to make sure

36

that you don't forget to connect to the virtual platform.

37

So let's do it.

38

I'm opening an icon the Navigator.

39

You have to find an X on the Navigator.

40

If you're on Windows you will find it in the list of programs and on Linux you can open it through either

41

to terminal or in the programs.

42

All right so now and I can the navigator is opened.

43

And don't forget to do this applications on virtual platform.

44

There we go.

45

Now we're connected to the virtual platform environment.

46

And so we were ready to launch spider.

47

And we don't have to install anything.

48

Everything is already installed on the virtual platform.

49

So we are ready to execute the code and I'm super happy to start.

50

But before we start implementing the code we have to be in the right folder because there are some external

51

files that we'll be calling when executing the code in the end.

52

So anyway we always have to be in the right folder where we implement the code.

53

So that's the first thing I'm going to do now I'm going to go to my desktop.

54

This is where my computer vision is at full that is.

55

So let's double click on it and now congratulations you reached module to object detection.

56

All right.

57

That's the folder.

58

Let's quickly describe what's inside this folder.

59

So you have data is just a folder that contains the classes based transform that will do the required

60

transformations so that the input images will be compatible with the neural network then.

61

Funny Doug is of course this video of this very funny.

62

We will be trying to detect.

63

I will show you this video in a second.

64

But then layer's is another folder that contained some other tools for the detection and the multi-book

65

as part of the SSD.

66

Then you have of course to code the commented version of the code object detection commented where you

67

have the whole code that will implemented this module come into line by line so that can be useful either

68

before or after.

69

Actually I also recommend to have a look at this before so that you can expect what you need to understand

70

and therefore when I explain it you might understand it more easily than this code object detection

71

is actually going to open it.

72

Is the code that we will implement in this module so I already imported the libraries.

73

I'm going to describe what those libraries are.

74

But anyway this is where I will implement this whole code and when I'm done implementing it with you

75

I will rename it object detection.

76

No comment that you can have the commented version of the code.

77

And the non-committed version of the code you can practice to recoated.

78

That's excellent practice.

79

Then we have the SSD that you wife file which contains the architecture of the single shot multi-button

80

action model.

81

We won't implement this one because I want it to keep what's most important for you to understand in

82

this object detection implementation.

83

Because if we implement the whole model this will be overwhelming.

84

And you might miss what's at the heart of the model.

85

So I prefer to proceed this way.

86

And this model is all the architecture.

87

And in fact after you watched in tuition lectures you will be totally able to understand what's going

88

on.

89

Well it's mostly about the architecture with all the boxes how they're defined.

90

But then the heart of the model will be in this implementation objective section.

91

And then finally this file is the file we will be loading to get the pre-trained SS DeMaio and more

92

precisely this is the file that contains the weight of the SSD neural network that was already pre-trained.

93

So we will be loading this file with torch the torch library that load which is a function of torche

94

this tortured load function will open a tensor a tensor that will contain the weight of this already

95

pre-trained neural network and then through a mapping with a dictionary we will transfer these weights

96

to the model we implement.

97

So basically this just contains the weight of an already pre-trained model and we will transfer these

98

weights to the model we will implement.

99

I hope that's clear and that's it.

100

So I guess we're ready to start.

101

And therefore let's start with some funny video of this very cute Doug.

102

So I'm going to double click on the video.

103

There we go.

104

That's the video you can recognize.

105

Kiril going to put that at the beginning.

106

So this is curial.

107

This is the dog.

108

This video last two seconds so that it doesn't take too much time.

109

When you try to Marans video.

110

But we will totally have time to see the dog bouncing.

111

It's very funny.

112

Check this out.

113

It you Doug when I watched this Doug I absolutely want to play with him.

114

And actually you can see Kyrle piloting the drone behind.

115

So there we go.

116

That's the video.

117

And actually the model we will implement will not only detect the dog bouncing on the field but also

118

this human here.

119

And you will see that it will also manage to detect curial even if you're really far actually from the

120

video.

121

And I'd like to tell you now that actually you know for you it's very easy to detect the drug but the

122

drug is actually pretty small in the video.

123

You know it's a pretty small object.

124

And when we tried to detect that with open city we had extremely bad results.

125

It couldn't detect the drug it couldn't detect what it was and there were some rectangles everywhere

126

you can actually try yourself.

127

And that's why I wanted to highlight that open Svea is definitely not among the most powerful models

128

but you'll see that the more we will implement in this module will do a perfect job at detecting this

129

drug even if it is small and even if there is not a perfect contrast between the drug and environment

130

you know it's not like we have a white environment with a black dog.

131

The dog can be confused with something else.

132

So you'll be convinced of the power of this model at the end and I can't wait to show you how this model

133

is going to do all right.

134

That's what I wanted to catch.

135

You know sometimes it really doesn't look like a dyke but you'll see what happens.

136

Let's implement the SSD single shot multi-book detection and let's do that from the next tutorial.

137

Until then enjoy computer vision.