1
00:00:00,000 --> 00:00:05,040
So how do you tell if your object detection algorithm is working well?

2
00:00:05,040 --> 00:00:10,234
In this video, you'll learn about a function called, "Intersection Over Union".

3
00:00:10,234 --> 00:00:14,115
And as we use both for evaluating your object detection algorithm,

4
00:00:14,115 --> 00:00:16,485
as well as in the next video,

5
00:00:16,485 --> 00:00:20,625
using it to add another component to your object detection algorithm,

6
00:00:20,625 --> 00:00:22,610
to make it work even better.

7
00:00:22,610 --> 00:00:25,495
Let's get started. In the object detection task,

8
00:00:25,495 --> 00:00:28,920
you expected to localize the object as well.

9
00:00:28,920 --> 00:00:31,870
So if that's the ground-truth bounding box,

10
00:00:31,870 --> 00:00:35,890
and if your algorithm outputs this bounding box in purple,

11
00:00:35,890 --> 00:00:38,900
is this a good outcome or a bad one?

12
00:00:38,900 --> 00:00:44,610
So what the intersection over union function does,

13
00:00:44,610 --> 00:00:53,650
or IoU does, is it computes the intersection over union of these two bounding boxes.

14
00:00:53,650 --> 00:00:59,195
So, the union of these two bounding boxes is this area,

15
00:00:59,195 --> 00:01:06,090
is really the area that is contained in either bounding boxes,

16
00:01:06,090 --> 00:01:11,580
whereas the intersection is this smaller region here.

17
00:01:11,580 --> 00:01:18,850
So what the intersection of a union does is it computes the size of the intersection.

18
00:01:18,850 --> 00:01:22,598
So that orange shaded area,

19
00:01:22,598 --> 00:01:27,520
and divided by the size of the union,

20
00:01:27,520 --> 00:01:30,430
which is that green shaded area.

21
00:01:30,430 --> 00:01:34,195
And by convention, the low compute division task will

22
00:01:34,195 --> 00:01:39,355
judge that your answer is correct if the IoU is greater than 0.5.

23
00:01:39,355 --> 00:01:45,310
And if the predicted and the ground-truth bounding boxes overlapped perfectly,

24
00:01:45,310 --> 00:01:47,054
the IoU would be one,

25
00:01:47,054 --> 00:01:50,105
because the intersection would equal to the union.

26
00:01:50,105 --> 00:01:55,195
But in general, so long as the IoU is greater than or equal to 0.5,

27
00:01:55,195 --> 00:01:59,685
then the answer will look okay, look pretty decent.

28
00:01:59,685 --> 00:02:03,880
And by convention, very often 0.5 is used as

29
00:02:03,880 --> 00:02:10,130
a threshold to judge as whether the predicted bounding box is correct or not.

30
00:02:10,130 --> 00:02:11,650
This is just a convention.

31
00:02:11,650 --> 00:02:12,975
If you want to be more stringent,

32
00:02:12,975 --> 00:02:14,790
you can judge an answer as correct,

33
00:02:14,790 --> 00:02:19,845
only if the IoU is greater than equal to 0.6 or some other number.

34
00:02:19,845 --> 00:02:21,570
But the higher the IoUs,

35
00:02:21,570 --> 00:02:24,425
the more accurate the bounding the box.

36
00:02:24,425 --> 00:02:27,625
And so, this is one way to map localization,

37
00:02:27,625 --> 00:02:32,560
to accuracy where you just count up the number of times an algorithm correctly

38
00:02:32,560 --> 00:02:37,815
detects and localizes an object where you could use a definition like this,

39
00:02:37,815 --> 00:02:42,410
of whether or not the object is correctly localized.

40
00:02:42,410 --> 00:02:46,515
And again 0.5 is just a human chosen convention.

41
00:02:46,515 --> 00:02:49,535
There's no particularly deep theoretical reason for it.

42
00:02:49,535 --> 00:02:54,640
You can also choose some other threshold like 0.6 if you want to be more stringent.

43
00:02:54,640 --> 00:03:00,070
I sometimes see people use more stringent criteria like 0.6 or maybe 0.7.

44
00:03:00,070 --> 00:03:04,100
I rarely see people drop the threshold below 0.5.

45
00:03:04,100 --> 00:03:08,065
Now, what motivates the definition of IoU,

46
00:03:08,065 --> 00:03:10,540
as a way to evaluate whether or not

47
00:03:10,540 --> 00:03:14,080
your object localization algorithm is accurate or not.

48
00:03:14,080 --> 00:03:20,340
But more generally, IoU is a measure of the overlap between two bounding boxes.

49
00:03:20,340 --> 00:03:22,430
Where if you have two boxes,

50
00:03:22,430 --> 00:03:23,980
you can compute the intersection,

51
00:03:23,980 --> 00:03:29,040
compute the union, and take the ratio of the two areas.

52
00:03:29,040 --> 00:03:34,985
And so this is also a way of measuring how similar two boxes are to each other.

53
00:03:34,985 --> 00:03:37,535
And we'll see this use again this way

54
00:03:37,535 --> 00:03:40,225
in the next video when we talk about non-max suppression.

55
00:03:40,225 --> 00:03:46,170
So that's it for IoU or Intersection over Union.

56
00:03:46,170 --> 00:03:50,720
Not to be confused with the promissory note concept in IoU,

57
00:03:50,720 --> 00:03:53,610
where if you lend someone money they write you a note that says,

58
00:03:53,610 --> 00:03:55,940
" Oh I owe you this much money," so that's also called an IoU.

59
00:03:55,940 --> 00:03:58,110
It's totally a different concept,

60
00:03:58,110 --> 00:04:03,111
that maybe it's cool that these two things have a similar name.

61
00:04:03,111 --> 00:04:07,730
So now, onto this definition of IoU, Intersection of Union.

62
00:04:07,730 --> 00:04:09,055
In the next video,

63
00:04:09,055 --> 00:04:12,045
I want to discuss with you non-max suppression,

64
00:04:12,045 --> 00:04:16,770
which is a tool you can use to make the outputs of YOLO work even better.

65
00:04:16,770 --> 00:04:18,470
So let's go on to the next video.