1
00:00:00,450 --> 00:00:08,220
And welcome back to Lesson 25, where we take a look at using Estes and open TV now essences and Sun

2
00:00:08,220 --> 00:00:13,950
for a single shot detectors which are a type of object detector that is quite similar to your which

3
00:00:13,950 --> 00:00:17,100
you used before, they're not as good as you all.

4
00:00:17,100 --> 00:00:22,710
In most cases, however, they are quite fast and they can be trained with a TensorFlow object detection

5
00:00:22,710 --> 00:00:26,790
library and a few at a library, so it's quite simple and easy to use sometimes.

6
00:00:27,240 --> 00:00:29,010
So let's look at our images here.

7
00:00:29,010 --> 00:00:34,870
Notebook systems do work quite well on the CPU's as well, which is one advantage they have over you.

8
00:00:35,370 --> 00:00:42,210
In my experience, using them on a video with open CVS, lower down and function here is designed and

9
00:00:42,210 --> 00:00:43,290
function functions.

10
00:00:43,900 --> 00:00:44,760
All right here.

11
00:00:45,780 --> 00:00:46,860
It works quite well.

12
00:00:46,890 --> 00:00:53,460
So the image size is quite small, though it resizes it to be 300 by 200 for the input for the processing.

13
00:00:53,940 --> 00:01:00,240
These are the hardcoded skill factors here, and this is a model we point to the model here that we

14
00:01:00,240 --> 00:01:00,720
downloaded.

15
00:01:00,960 --> 00:01:02,910
So you can see it here.

16
00:01:03,180 --> 00:01:09,060
This is a model fails the inference graph as well as the Peavey text, which does take a look at that

17
00:01:09,060 --> 00:01:10,320
and see what it is.

18
00:01:10,830 --> 00:01:12,990
This here is the model definition.

19
00:01:13,110 --> 00:01:17,120
This is a bit bigger, so it's similar to the euro.

20
00:01:17,130 --> 00:01:23,550
See if she follows just the model description so that when we load, it knows exactly what is what.

21
00:01:24,150 --> 00:01:28,990
So these are things you would learn in the deep learning section of this course.

22
00:01:29,040 --> 00:01:36,090
So for now, to worry about it, this model has been trained on 90 classes, one of the Cooku data sets

23
00:01:36,090 --> 00:01:37,140
with 90 classes.

24
00:01:37,140 --> 00:01:38,670
I believe so.

25
00:01:38,670 --> 00:01:40,970
Yeah, mobile at Koko.

26
00:01:40,980 --> 00:01:41,640
This one here.

27
00:01:42,270 --> 00:01:46,800
So actually, this SSD is a mobile on that backbone SSD.

28
00:01:47,250 --> 00:01:49,330
That's the one we're using here, which is quite fast.

29
00:01:49,350 --> 00:01:54,180
It's one of the few models that works at maybe 10 frames a second on most cell phones.

30
00:01:55,680 --> 00:01:57,750
So sorry for scrolling.

31
00:01:57,750 --> 00:01:58,710
So these here.

32
00:01:59,700 --> 00:02:05,790
So when you look the model, we defined class names here, so you can see all the costumes of the.

33
00:02:06,350 --> 00:02:07,080
We have a dictionary.

34
00:02:07,080 --> 00:02:13,110
It looks up the class name from the number we can image here and it goes into frame.

35
00:02:14,370 --> 00:02:20,760
Then we just followed passage through the network and then we just get the rose cones.

36
00:02:21,240 --> 00:02:22,410
So we get a square input.

37
00:02:22,410 --> 00:02:26,250
So we just get the square center of the image that we pass.

38
00:02:27,510 --> 00:02:29,460
And then we just plot the points here.

39
00:02:29,970 --> 00:02:35,700
So similarly, like in the yellow one, we just get the detections out of it here and we just get the

40
00:02:35,700 --> 00:02:42,210
confidence of the confidence is than a only to find teacher was defined to be at point five in this

41
00:02:42,210 --> 00:02:42,600
case.

42
00:02:43,320 --> 00:02:44,420
And we just loop through.

43
00:02:44,430 --> 00:02:49,260
We do want to play non maximal suppression and mass in this one.

44
00:02:49,830 --> 00:02:51,150
So let's take a look at the outputs.

45
00:02:51,150 --> 00:02:52,140
Let's run this code.

46
00:02:54,630 --> 00:02:58,230
It should be quite fast once we learn all the models in the model.

47
00:02:58,620 --> 00:03:04,170
Sometimes it takes, sometimes takes the longest amount of time, and inference afterwards is just in

48
00:03:04,170 --> 00:03:04,920
the milliseconds.

49
00:03:05,640 --> 00:03:07,230
So this is the original image here.

50
00:03:07,800 --> 00:03:12,840
These are the cops here, so you can definitely see it's not performing as good as usual, and the boxes

51
00:03:12,840 --> 00:03:14,370
do have some weird overlap.

52
00:03:14,850 --> 00:03:17,280
This so that's that.

53
00:03:17,620 --> 00:03:19,590
That's a that's it for that image.

54
00:03:20,010 --> 00:03:21,670
We can try out the images here.

55
00:03:21,720 --> 00:03:28,500
Let's try another one, the shredded elephant, and see how that looks.

56
00:03:30,000 --> 00:03:35,340
So again, we give it some time to run usually takes about 10 seconds.

57
00:03:35,700 --> 00:03:37,520
Most of that time is spent in the models.

58
00:03:37,540 --> 00:03:42,450
If you wanted, you can just loop through the model once and then run the inferences on the images afterward,

59
00:03:42,630 --> 00:03:43,680
like what we did for you.

60
00:03:44,700 --> 00:03:45,190
There we go.

61
00:03:45,210 --> 00:03:50,010
So you can see it actually worked quite well with the elephant and high confidence that it is an elephant

62
00:03:50,730 --> 00:03:52,920
so you can experiment with Estes.

63
00:03:53,160 --> 00:04:00,780
As I said, I would use the yellow in most cases for other object vectors of colors, people, animals

64
00:04:01,290 --> 00:04:10,950
or any of the Cooku Objects SSD, as I would use an SSD like a mobile net backbone SSD when the system

65
00:04:10,950 --> 00:04:16,080
I'm running is on like a Raspberry Pi or some embedded system where inference speed.

66
00:04:16,530 --> 00:04:17,400
Well, the competition.

67
00:04:17,430 --> 00:04:22,710
You want it to be quite fast and efficient because that that was embedded systems on powerful.

68
00:04:23,310 --> 00:04:24,600
So you would I would use those.

69
00:04:25,080 --> 00:04:26,640
So thank you for that lesson.

70
00:04:27,120 --> 00:04:33,930
I will now move on to colorizing black and white photos using a cafe model in Open City.

71
00:04:34,860 --> 00:04:35,850
So stay tuned for that.

72
00:04:36,030 --> 00:04:36,420
Thank you.