1
00:00:00,540 --> 00:00:05,460
Hi and welcome to your 15 to open TV chapter, which is called Finding Corners.

2
00:00:05,910 --> 00:00:10,410
So go ahead and open this notebook and you can see the score up to the top.

3
00:00:11,160 --> 00:00:15,300
Firstly, let's load all images and libraries and everything.

4
00:00:15,960 --> 00:00:16,440
The function?

5
00:00:17,160 --> 00:00:19,300
So let's go down to this sufficiently.

6
00:00:19,320 --> 00:00:20,280
What is a corner?

7
00:00:20,550 --> 00:00:22,720
Because I mean, intuitively, you know a corner.

8
00:00:22,770 --> 00:00:24,480
What a corner is an image.

9
00:00:25,170 --> 00:00:25,800
I mean, you should.

10
00:00:26,400 --> 00:00:28,230
However, I'll explain to you a bit.

11
00:00:28,530 --> 00:00:32,610
Corner is essentially it's it's a local point.

12
00:00:32,940 --> 00:00:38,460
Remember, it's corner to corner point, whose local neighbourhood stands in two dominant and different

13
00:00:38,460 --> 00:00:44,490
edged directions, which means that it's basically the threshold point for a big change in the image.

14
00:00:45,060 --> 00:00:50,730
And in other words, a corner can be interpreted as a junction which have two edges where edges is such

15
00:00:51,030 --> 00:00:53,010
a sudden change in the image brightness.

16
00:00:53,760 --> 00:00:59,330
Corners are important features and an image, and generally a team pinned to it as an interesting point,

17
00:00:59,340 --> 00:01:05,280
and this is an important line is that they're invariant to translations and rotation and illumination,

18
00:01:05,280 --> 00:01:07,470
which means that they're very good key points.

19
00:01:08,040 --> 00:01:15,030
So one of the best quality detectors, which is Harris Quinn Interactive, which was developed in 1998.

20
00:01:15,720 --> 00:01:16,830
You can read the paper here.

21
00:01:17,280 --> 00:01:19,040
It's a quite a quite a good algorithm.

22
00:01:19,050 --> 00:01:21,580
I'm not going to go through the entire algorithm.

23
00:01:21,600 --> 00:01:25,980
I'm just going to go through how to use the algorithm in open TV, but it works quite well.

24
00:01:26,580 --> 00:01:29,310
So you can see the actually output of this chessboard pattern here.

25
00:01:29,340 --> 00:01:31,230
It actually gets all the corners fairly well.

26
00:01:31,590 --> 00:01:38,100
And I've used this for two point extraction and other images, and they generally generally works very,

27
00:01:38,100 --> 00:01:38,670
very well.

28
00:01:38,880 --> 00:01:40,560
You don't really have much issues with this.

29
00:01:40,560 --> 00:01:43,920
And even with these default parameters, it works quite well.

30
00:01:44,460 --> 00:01:46,020
So let's take a look at the function.

31
00:01:46,110 --> 00:01:52,290
If this is the function here key to coin a house and all of those, it takes a grayscale image as the

32
00:01:52,290 --> 00:01:54,540
input and we use the block size.

33
00:01:54,540 --> 00:01:58,920
It's a neighborhood that we're going to use to consider neighbourhood as pixels around each area.

34
00:01:58,920 --> 00:02:06,990
So then to give it a sliding window, a type size we're looking at and then we look at aperture size,

35
00:02:07,350 --> 00:02:11,130
which is what we use it or we use to calculate a sort of derivative.

36
00:02:11,730 --> 00:02:14,370
And then there's the key, which is the Harris detector.

37
00:02:14,520 --> 00:02:19,020
It's a free parameter in this equation, which I'm not entirely sure what it controls.

38
00:02:19,050 --> 00:02:23,990
Maybe you can experiment and see for yourself or read the paper and see if you can figure it out.

39
00:02:24,000 --> 00:02:28,470
And actually, this this algorithm was about in 1980, not 1938.

40
00:02:29,010 --> 00:02:30,210
That's my my table.

41
00:02:30,780 --> 00:02:32,210
So let's just correct that.

42
00:02:32,220 --> 00:02:37,680
So let's run this code and you can see the outputs of this, and it looks looks quite good.

43
00:02:38,550 --> 00:02:43,920
So there's a little corner detection algorithm, which is called good for just the traffic, which is

44
00:02:44,100 --> 00:02:45,180
quite explicit.

45
00:02:45,180 --> 00:02:49,620
And what it is gives us good features and also it also gives us good corners.

46
00:02:50,160 --> 00:02:54,120
Perhaps I haven't used this much in practice, but this is the output here.

47
00:02:54,630 --> 00:02:58,940
It's quite simple again, and you have the best configuration parameters.

48
00:02:59,010 --> 00:03:01,530
It tells us how much corners we want to look at.

49
00:03:02,610 --> 00:03:03,710
What's a quality level?

50
00:03:03,720 --> 00:03:10,440
Quality levels is basically how strict, how on strict we want to be with what a corner is, how we

51
00:03:10,440 --> 00:03:14,790
define a corner essentially, and the minimum distance between returned corners.

52
00:03:15,210 --> 00:03:23,670
So let's run this block of code here and you can see we take the output outputs of basically x y points

53
00:03:23,670 --> 00:03:26,690
and we can just draw rectangles around those points.

54
00:03:26,700 --> 00:03:33,120
So we just go ten pixels above 10 pixels wide and destroy these little rectangles around here.

55
00:03:33,570 --> 00:03:37,930
So those are two very interesting corner detection methods in open TV.

56
00:03:39,000 --> 00:03:44,130
You may have to use it or not for your applications, depending on what it is, but it's very good to

57
00:03:44,130 --> 00:03:44,610
know this.

58
00:03:45,000 --> 00:03:48,480
So that concludes, most of all, a simple, open TV work.

59
00:03:49,350 --> 00:03:53,220
These are all image manipulations, not all of them.

60
00:03:53,220 --> 00:03:54,570
They're actually quite a few more.

61
00:03:54,930 --> 00:03:56,520
But these are the most useful ones.

62
00:03:56,520 --> 00:04:02,190
I find that actually as a beginner practitioner, the ones you need to know are all here in this first

63
00:04:02,190 --> 00:04:03,030
15 chapters.

64
00:04:03,540 --> 00:04:09,330
And now we're going to move on to a bit more advanced settings, which is fierce and eye detection and

65
00:04:09,330 --> 00:04:10,680
a bunch of other algorithms here.

66
00:04:11,070 --> 00:04:17,390
So stay tuned, and we'll be doing a lot more fun and cool stuff with computer vision and open TV.

67
00:04:17,550 --> 00:04:18,000
Thank you.