5. OpenCV Tutorial - Learn Classic Computer Vision & Face Detection (OPTIONAL)/27. Approximating Contours & Finding Their Convex Hull - Clean Up Messy Contours.srt

1
00:00:00,660 --> 00:00:04,120
So what we're going to look at now is called approximating contours.

2
00:00:04,480 --> 00:00:06,700
So fleecy let's open this file here.

3
00:00:06,750 --> 00:00:08,310
This is 4.3.

4
00:00:08,700 --> 00:00:09,530
4.3 here

5
00:00:13,300 --> 00:00:16,060
and that's actually run this could see what's going on.

6
00:00:16,420 --> 00:00:20,680
So I have a very crudely drawn house that I did by hand to do a splint.

7
00:00:20,920 --> 00:00:22,550
Thank you very much.

8
00:00:22,660 --> 00:00:26,330
And this is what happens when we extract contours and approx.

9
00:00:26,340 --> 00:00:31,360
I used a bonding rectangle for Contos so you can see we've have we have a bonding rectangle for the

10
00:00:31,360 --> 00:00:32,530
entire house.

11
00:00:32,680 --> 00:00:33,810
Then one for the roof.

12
00:00:33,820 --> 00:00:39,460
So I'm guessing find Quanto has found contours of this triangle here than the door and then the bottom

13
00:00:39,460 --> 00:00:40,560
half of the hose here.

14
00:00:40,810 --> 00:00:42,130
So that's pretty cool.

15
00:00:42,170 --> 00:00:45,720
Now there they're prox to approx function is what we're going to look at.

16
00:00:45,720 --> 00:00:46,300
No.

17
00:00:46,650 --> 00:00:52,660
And what does that actually as you can see here trace approximate or contour shape better.

18
00:00:53,140 --> 00:01:00,770
So this may look a lot similar to the actual initial contour operation here.

19
00:01:00,900 --> 00:01:04,450
However what it does it actually tries to approximate a polygon here.

20
00:01:04,500 --> 00:01:06,060
So let's take a look at a function here.

21
00:01:10,190 --> 00:01:11,660
So you can see you know a loop here.

22
00:01:11,660 --> 00:01:12,490
We're actually doing.

23
00:01:12,530 --> 00:01:14,030
We're looking through the Contos again.

24
00:01:14,030 --> 00:01:15,950
We found find Contos right here.

25
00:01:16,160 --> 00:01:19,650
And refining all contours and doing no approximation here.

26
00:01:19,730 --> 00:01:24,380
Maybe I should have done a simple approximation although it wouldn't have made much difference in this

27
00:01:24,380 --> 00:01:25,190
example here.

28
00:01:28,110 --> 00:01:35,700
Going back to Otari approximate poly DP function here in TV2 This is how we approximate Contos is actually

29
00:01:35,700 --> 00:01:41,430
very useful if you have a noisy image with jagged contours and you know a ship is supposed to be square

30
00:01:41,490 --> 00:01:47,880
or a triangle whoever you're getting maybe multiple edges and that polygon dysfunction actually basically

31
00:01:48,770 --> 00:01:51,020
approximates approximates it for you.

32
00:01:51,030 --> 00:01:54,390
So let's look at the parameters it takes in here.

33
00:01:54,410 --> 00:01:55,410
So at a high level.

34
00:01:55,430 --> 00:01:57,500
Describe exactly what his function is doing.

35
00:01:57,710 --> 00:02:04,280
So it takes tree parameters into consideration here it takes input CONTO something called an approximation

36
00:02:04,310 --> 00:02:06,520
accuracy which I'll discuss shortly.

37
00:02:06,650 --> 00:02:11,570
And basically what do we want on a closed polygon or open polygon.

38
00:02:11,640 --> 00:02:15,080
So is being preferred in pretty much all cases I believe.

39
00:02:15,200 --> 00:02:21,020
So let's take a look at approximation accuracy approximation accuracy is basically how detailed the

40
00:02:21,020 --> 00:02:25,910
whole finally agree and do you want the approximation to be if you want to low accuracy and a good rule

41
00:02:25,910 --> 00:02:28,960
of thumb here is 5 percent of the contour perimeter.

42
00:02:29,000 --> 00:02:35,660
However if you go to high efficiency with accuracy let's look at c using Point 1 here.

43
00:02:35,660 --> 00:02:40,870
Let's see what happens it actually does something pretty weird it actually tries to approximate the

44
00:02:40,870 --> 00:02:44,350
house into the freedom of the house into a triangle.

45
00:02:44,350 --> 00:02:45,670
That's clearly not good.

46
00:02:45,940 --> 00:02:46,450
So what if.

47
00:02:46,470 --> 00:02:53,820
What about if we use point 0 1 as you can see it actually tries to match the house even closer.

48
00:02:53,830 --> 00:02:58,810
So what I suggest is that you maybe maybe play with these values a bit when he actually works fairly

49
00:02:58,810 --> 00:03:00,700
well in this case here.

50
00:03:00,940 --> 00:03:06,580
Basically the accuracy or to reassure to perimeter the more precise the approximation is going to be

51
00:03:07,030 --> 00:03:13,410
and the higher value here the more logic green that the approximation is going to be.

52
00:03:13,660 --> 00:03:17,310
So let's keep that in mind when you're actually playing with the accuracy parameter here.

53
00:03:18,640 --> 00:03:23,440
OK so let's move on to the convex hull and the convex hole is an interesting concept and it's actually

54
00:03:23,440 --> 00:03:27,070
run Decoud to illustrate this.

55
00:03:27,200 --> 00:03:31,110
So as you can see we have a hand here sort of like a blue.

56
00:03:31,260 --> 00:03:37,900
And right now and continuing we have what is called the convex hull not a convex hole.

57
00:03:37,910 --> 00:03:43,670
Basically if you can kind of figured out from this image here it looks a different do outer edges here

58
00:03:43,760 --> 00:03:46,590
and it draws lines to each edges here.

59
00:03:46,610 --> 00:03:54,410
Basically it's what you call the smallest polygon that can fit or the object itself and HUD's found

60
00:03:54,440 --> 00:03:56,410
it actually is pretty simple as well.

61
00:03:58,260 --> 00:04:04,140
So all that's happening here is that in this loop here we're digging contours that we found here.

62
00:04:04,560 --> 00:04:06,660
There's actually two ends of quid I'll explain shortly here.

63
00:04:06,660 --> 00:04:08,570
It's not too critical just something it does.

64
00:04:08,610 --> 00:04:08,870
I did.

65
00:04:08,870 --> 00:04:16,850
And just to show you guys and using the CV to convex whole function it generates a whole data type here

66
00:04:16,920 --> 00:04:18,320
which is actually in tree.

67
00:04:18,600 --> 00:04:24,540
And then we put the saree into distro Kontos which takes in a list function a list variable I should

68
00:04:24,540 --> 00:04:30,180
say and we just draw all the contorts off the image.

69
00:04:30,530 --> 00:04:32,680
So that's all there is to it and convex whole.

70
00:04:32,960 --> 00:04:39,200
It's actually quite simple to implement and use this sort of line of code is actually when you use to

71
00:04:39,380 --> 00:04:42,170
retrieve list method what happens is that.

72
00:04:42,200 --> 00:04:44,730
Oh mislaying let's just keep this.

73
00:04:44,770 --> 00:04:45,470
Whoops.

74
00:04:48,130 --> 00:04:49,130
And do that.

75
00:04:49,230 --> 00:04:51,650
Let's just keep this here the same.

76
00:04:51,690 --> 00:04:56,430
So let's run the code again and you'll see a box drawn over this image here.

77
00:04:56,700 --> 00:05:02,590
And what happens is that every time you find a console and a white image here it actually identifies

78
00:05:02,590 --> 00:05:05,530
the white background itself as a conto.

79
00:05:05,610 --> 00:05:09,260
If it was a black image it wouldn't have happened in this way.

80
00:05:09,390 --> 00:05:17,070
So to get rid of that which happens what I do I tend to do sometimes just use this little lines of code

81
00:05:17,070 --> 00:05:22,890
here that just if I wanted all Kontos but I wanted to eliminate that largest CONTO I saw the Contos

82
00:05:22,890 --> 00:05:30,930
Bay Area and they just use non-pay to index it at everything except the first or largest Hanzal and

83
00:05:30,930 --> 00:05:32,460
that's simply what they do here.

84
00:05:32,610 --> 00:05:39,340
So this is a way to get rid of that first box or an entire image and generally just decline to be where

85
00:05:39,370 --> 00:05:40,950
I want to look at here.