Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,770 --> 00:00:05,720 Hello and welcome back to the course on computer vision and today we're talking about training classifiers. 2 00:00:05,720 --> 00:00:10,760 So previously we discussed that Develin Jones algorithm has two stages training and detection. 3 00:00:10,790 --> 00:00:14,870 We've already spoken about detection and today's going to be a quick tutorial to get us started into 4 00:00:14,870 --> 00:00:16,370 the world of training. 5 00:00:16,820 --> 00:00:19,280 So what exactly do we mean when we say training. 6 00:00:19,460 --> 00:00:21,890 Well we're training these features as you remember. 7 00:00:21,890 --> 00:00:27,350 We've got five basic features but the trick is that they're scalable they can be shorter or longer wider 8 00:00:27,410 --> 00:00:28,570 or narrower. 9 00:00:28,730 --> 00:00:38,570 And we need to identify which ones of them are actually descriptive or a face or something are common 10 00:00:38,570 --> 00:00:40,020 features of a face. 11 00:00:40,040 --> 00:00:42,130 So here we've got a face. 12 00:00:42,210 --> 00:00:49,100 And remember we talked about that here the bridge of the nose is lighter and this is darker well. 13 00:00:49,220 --> 00:00:56,030 We know that from our experience with faces but an algorithm because it's probably the first face that 14 00:00:56,030 --> 00:00:57,170 it's seen. 15 00:00:57,290 --> 00:01:02,180 Won't necessarily know that that is a common feature for faces for instance from this photo It might 16 00:01:02,180 --> 00:01:02,970 think that. 17 00:01:03,180 --> 00:01:08,420 OK so when this part is lighter and this part is darker that's also a common feature for faces but that's 18 00:01:08,420 --> 00:01:11,230 not necessarily true because not everybody has a beard. 19 00:01:11,510 --> 00:01:15,610 On the other hand it might also look at the background might think that you know this white and this 20 00:01:15,620 --> 00:01:20,510 black this this contrast is also called Faces which is also not the case. 21 00:01:20,510 --> 00:01:27,680 So we need to help the algorithm understand which features are important to identify a face which features 22 00:01:27,680 --> 00:01:28,760 are not important. 23 00:01:28,820 --> 00:01:34,520 So that's number one and number two is remember when we're talking about features we spoke about the 24 00:01:34,520 --> 00:01:39,210 threshold that you will never get like exactly. 25 00:01:39,230 --> 00:01:47,030 Very bright white here and very dark black here it all is a form of grey greyscale So a bit maybe of 26 00:01:47,300 --> 00:01:54,290 some contrast like not exactly 100 percent dark but close to dark and close to bright. 27 00:01:54,320 --> 00:02:00,680 So there's some there's some difference in between the contrast and that difference has to be it has 28 00:02:00,680 --> 00:02:08,300 to exceed the threshold for us to consider this feature to be present while the training will also help 29 00:02:08,300 --> 00:02:11,500 the algorithm understand those thresholds should they be set up 0.3. 30 00:02:11,570 --> 00:02:17,780 Should they be set at zero 0.7 position to be sets and zero point fifty seven or 56 or something like 31 00:02:18,110 --> 00:02:21,590 it needs to understand where to set those thresholds for the different features. 32 00:02:21,590 --> 00:02:22,500 And that's part of the training. 33 00:02:22,500 --> 00:02:28,700 So first of all identify the features and second to set thresholds and that's that's in a very brief 34 00:02:28,700 --> 00:02:30,380 overview of what the training is about. 35 00:02:30,380 --> 00:02:37,520 We'll talk more about it in coming to terms but basically that's kind of the goal to understand what 36 00:02:37,640 --> 00:02:44,540 is what features identify a face and how to say how to know when their presence or basically the threshold 37 00:02:44,540 --> 00:02:45,460 side of things. 38 00:02:45,560 --> 00:02:54,820 So what the algorithm does is it shrinks the image to 24 by 24 pixels and then it looks for these different 39 00:02:54,820 --> 00:03:00,070 features on the image it tries to understand which features are common for faces so the first question 40 00:03:00,070 --> 00:03:02,190 is Why does a trinket 24 by 24. 41 00:03:02,410 --> 00:03:05,200 Well because as we discussed these features are scalable. 42 00:03:05,200 --> 00:03:12,610 So when you have a very large image of for instance a thousand by a thousand or even 500 by 500 pixels 43 00:03:13,090 --> 00:03:19,940 then there's lots of different variations of these features so it can be like one pixel by one pixel 44 00:03:19,940 --> 00:03:25,930 that can be two by two it can be so like two there too there can be four pixels in a four pixel that 45 00:03:25,930 --> 00:03:31,780 would be 10 and 10 would be 100 100 so there's just lots and lots of different combinations of these 46 00:03:31,780 --> 00:03:38,730 features are a combination different lengths and widths and heights of these features. 47 00:03:38,770 --> 00:03:43,440 And you just take forever to look through them and understand which ones are telling her face. 48 00:03:43,630 --> 00:03:50,200 So it's easier to scale down the image and then apply a very limited and we'll talk about this for the 49 00:03:50,200 --> 00:03:56,570 down as well and apply the features that fit in a 24 hour 24 pixel window so it's much less combinations. 50 00:03:56,800 --> 00:04:02,680 And then once we found them once we found OK so that you will still find here that the nose is like 51 00:04:02,950 --> 00:04:09,960 brighter and darker and you'll still find the same features but then when you're actually applying it 52 00:04:09,960 --> 00:04:16,140 to a Nimish to Detective face what happens then is we apply we keep the image the same size we don't 53 00:04:16,140 --> 00:04:20,080 scale it down to 24 by 24 but then we scale the features up. 54 00:04:20,220 --> 00:04:20,910 So that's the trick. 55 00:04:20,910 --> 00:04:23,820 So when you're training you scale the image down. 56 00:04:23,820 --> 00:04:27,360 All of that and we'll talk about many training images just now. 57 00:04:27,360 --> 00:04:33,990 But all of the training images are scaled down to 24 pixels and the features I detected there and the 58 00:04:33,990 --> 00:04:39,670 training happens there so we know which features are important and what what threshold they should have. 59 00:04:39,900 --> 00:04:44,940 And then in the real life when you're actually detecting you don't scale the image you keep the image 60 00:04:44,940 --> 00:04:48,440 at its original size wherever it is you scale the features up. 61 00:04:48,450 --> 00:04:52,350 So those features that are important you scale them up and then you look for them on the image. 62 00:04:52,410 --> 00:04:54,440 So that's an important thing to remember. 63 00:04:54,450 --> 00:04:55,990 That's the important trick there. 64 00:04:56,340 --> 00:04:57,950 And there you go. 65 00:04:57,990 --> 00:05:00,650 So what's what's missing here. 66 00:05:00,660 --> 00:05:02,200 We already know kind of the concept. 67 00:05:02,200 --> 00:05:06,780 Yeah we need to look for these features on a story for which we for image that we all understand that 68 00:05:07,110 --> 00:05:09,310 awaits what is missing is the data. 69 00:05:09,330 --> 00:05:10,750 One image is not enough. 70 00:05:10,770 --> 00:05:15,540 As we said it might pick up the wrong things from this one image. 71 00:05:15,780 --> 00:05:25,030 That's why you need to supply lots of images lots of faces of people frontal faces of people to the 72 00:05:25,030 --> 00:05:30,820 algorithms so that it can understand which of those features that it's looking for are actually common. 73 00:05:30,910 --> 00:05:33,820 So when it has lots of faces it'll be able to tell. 74 00:05:33,820 --> 00:05:39,000 OK so you know this feature that I saw on one face is repeating is happening on many many many many 75 00:05:39,000 --> 00:05:39,940 different faces. 76 00:05:40,090 --> 00:05:47,470 And so with you know like I can if I look for that feature then I'm going to have a good chance of finding 77 00:05:47,470 --> 00:05:53,830 a face whereas this other feature which I saw in one image is actually not present on most of the other 78 00:05:53,830 --> 00:05:55,750 images so it's not important to. 79 00:05:56,230 --> 00:06:03,570 And here we've only got a couple of faces but in the real algorithm in the actual original paper by 80 00:06:03,630 --> 00:06:09,940 Alan Jones supplied their algorithm four thousand nine hundred and sixteen manually labeled faces you 81 00:06:09,940 --> 00:06:15,970 can imagine how much work that was to go through four thousand nine hundred and sixteen photos and manual 82 00:06:15,970 --> 00:06:22,370 label that like each one instance a face is a face to face you know cut those faces out make them twenty 83 00:06:22,370 --> 00:06:23,400 four four. 84 00:06:23,650 --> 00:06:26,370 Make sure that there are faces in there and so on. 85 00:06:26,640 --> 00:06:30,100 And then they apply a quick trick solecism. 86 00:06:30,100 --> 00:06:34,970 If you're going to be training your own algorithm. 87 00:06:34,990 --> 00:06:36,630 So this is a very interesting trick. 88 00:06:36,640 --> 00:06:42,040 You just take forever for all the faces you found Take a mirror image of the face so left to right mirror 89 00:06:42,040 --> 00:06:42,760 image. 90 00:06:42,760 --> 00:06:46,410 So for instance here we've got a face where the person is more on the left in the mirror image should 91 00:06:46,420 --> 00:06:47,410 be a bit more to the right. 92 00:06:47,410 --> 00:06:51,640 So it would be just like his left eye would look like his right eye his right eye would look like his 93 00:06:51,640 --> 00:06:52,190 left eye. 94 00:06:52,330 --> 00:06:58,540 So it would be just a mirror image of this image and to us it would look pretty much the same you'd 95 00:06:58,540 --> 00:06:59,470 see that it's a mirror image. 96 00:06:59,480 --> 00:07:01,660 But for a computer it's a brand new image. 97 00:07:01,660 --> 00:07:07,060 It's technically a brand new image and therefore they instantly by doing that they doubled the number 98 00:07:07,060 --> 00:07:08,400 of images too. 99 00:07:08,410 --> 00:07:10,780 Nine hundred and nine thousand eight hundred thirty two. 100 00:07:11,200 --> 00:07:12,130 So that's good. 101 00:07:12,130 --> 00:07:19,630 So from here we can find all these 9000 almost 10000 images we can find which features are common for 102 00:07:19,630 --> 00:07:20,580 faces. 103 00:07:20,680 --> 00:07:22,000 Good and we just can keep those. 104 00:07:22,120 --> 00:07:28,900 But then we also need to make sure that those features are common just for faces that they're not common 105 00:07:28,900 --> 00:07:29,910 for something else. 106 00:07:30,130 --> 00:07:36,430 And that's why you need to supply also non face images needs a whole set of non-fatal images just some 107 00:07:36,420 --> 00:07:38,820 random photos objects everything. 108 00:07:38,830 --> 00:07:46,210 But you need to make sure that not of those images have faces in them and Viola Jones supplied their 109 00:07:46,210 --> 00:07:46,690 algorithm. 110 00:07:46,700 --> 00:07:51,240 Nine thousand five hundred and forty four of these images and. 111 00:07:51,420 --> 00:07:57,440 But the interesting thing here is that these ones don't have to be 24 or 24 because they're going to 112 00:07:57,440 --> 00:08:00,480 be any size and that of course is a limit. 113 00:08:00,480 --> 00:08:01,860 A couple of hundred pixels. 114 00:08:02,040 --> 00:08:09,300 But basically then even if it's a big image you can just take some windows from this image and treat 115 00:08:09,390 --> 00:08:15,600 each one as a separate image for training purposes and that increases even further the amount of data 116 00:08:15,600 --> 00:08:16,910 that you have for non-ferrous images. 117 00:08:16,920 --> 00:08:21,630 And in the old John's argument was about three hundred and fifty million sub windows that they had for 118 00:08:21,630 --> 00:08:22,480 training. 119 00:08:22,980 --> 00:08:26,410 So there you go as you can imagine together this will do the trick. 120 00:08:26,490 --> 00:08:31,740 When you have the face images there will help them understand which features are important for faces 121 00:08:31,770 --> 00:08:33,010 and we'll just keep those. 122 00:08:33,120 --> 00:08:38,460 And then the non face images will help it reconcile which of those features that it found that are good 123 00:08:38,460 --> 00:08:43,110 for faces are also leading to a higher rate of false positives. 124 00:08:43,110 --> 00:08:48,210 So for instance it found a certain feature it helps it pick up a face but it also helps it always pick 125 00:08:48,210 --> 00:08:48,870 up. 126 00:08:48,930 --> 00:08:50,020 I don't know a dog. 127 00:08:50,340 --> 00:08:52,040 And it just keeps picking up these dogs. 128 00:08:52,050 --> 00:08:58,530 But because these images are labeled as non faces it will be able to learn that that feature although 129 00:08:58,550 --> 00:09:02,940 is good for faces it's bad in the sense that it's picking up a lot of dogs. 130 00:09:02,970 --> 00:09:08,700 So I'm not going to keep that feature and that's how it'll work so it will use the face images which 131 00:09:08,700 --> 00:09:12,090 I was able to pick up to just limit the number of rooms. 132 00:09:12,090 --> 00:09:19,790 Let's go back room there are these these features that potentially fit into this window. 133 00:09:20,190 --> 00:09:25,450 It all of them will pick out the ones that are good for faces and then using the nonfatal images. 134 00:09:25,560 --> 00:09:32,310 It will drop the ones that are giving it high rates of false positives so that you can you can pick 135 00:09:32,310 --> 00:09:35,510 out faces and only the faces and nothing else. 136 00:09:35,820 --> 00:09:42,420 So that's in a nutshell how the training works and that helps pick out the features. 137 00:09:42,420 --> 00:09:47,760 And it also helps understand the thresholds that those features should have. 138 00:09:47,970 --> 00:09:55,530 If you'd like to do some additional reading on this topic then a petition good paper is a general framework 139 00:09:55,530 --> 00:09:59,450 for object detection by Constantine Puppa Giorgio. 140 00:09:59,700 --> 00:10:08,880 And it was written before surgery in 1998 before the paper by VELLAR Jones and that's where the horror 141 00:10:08,880 --> 00:10:11,370 like features were introduced in the first place. 142 00:10:11,730 --> 00:10:12,860 So you might want to check that out. 143 00:10:12,870 --> 00:10:19,990 But other other and that's of course the original VELLAR Jones paper has everything you need. 144 00:10:19,980 --> 00:10:24,450 It has all the details about the Horlick features and how everything works. 145 00:10:24,470 --> 00:10:34,560 Like I'll admit I haven't read in detail this paper or this new paper but like I looked at the abstract 146 00:10:34,620 --> 00:10:40,800 it looks interesting but personally I prefer the village on paper I think that's a full enough resource. 147 00:10:40,800 --> 00:10:44,070 So if you haven't read that one yet I highly recommend checking that one out. 148 00:10:44,310 --> 00:10:46,760 And on that note I look forward to an extent. 149 00:10:46,920 --> 00:10:48,890 And until then enjoy computer vision. 16730