Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,610 --> 00:00:06,010 Hello welcome back to the course on computer vision in today's tutorial we're going to talk about adaptive 2 00:00:06,160 --> 00:00:08,390 boosting or other boost. 3 00:00:08,440 --> 00:00:17,320 So previously we left off when we discussed how the algorithm uses input data or images to find which 4 00:00:17,320 --> 00:00:21,310 features are important to recognizing faces. 5 00:00:21,310 --> 00:00:28,610 So it uses the face images to understand which features are common among faces and it keeps those. 6 00:00:28,700 --> 00:00:34,040 And then it uses the nonfatal images to understand out of the ones that are the features that it's kept 7 00:00:34,460 --> 00:00:38,760 which features give it high rates of false positives. 8 00:00:38,930 --> 00:00:43,670 And then it discards those features and makes them less important and focus on the features that are 9 00:00:43,730 --> 00:00:51,830 present and faces and faces only that do not come up very commonly in other types of objects or the 10 00:00:51,860 --> 00:00:53,000 pictures. 11 00:00:53,070 --> 00:00:55,300 And so that's basically how the algorithm is trained. 12 00:00:55,640 --> 00:01:02,330 And then we have these different features and there are thresholds at which they are considered to be 13 00:01:02,330 --> 00:01:07,590 present in an image and so that sounds like the end of the story sounds like. 14 00:01:07,600 --> 00:01:08,180 Ok cool. 15 00:01:08,180 --> 00:01:13,700 We've identified the features we've identified thresholds all we have to do when a new image which is 16 00:01:13,730 --> 00:01:19,870 going to watch is going to look at the features that we've identified check if their present check the 17 00:01:19,910 --> 00:01:20,770 thresholds are met. 18 00:01:20,780 --> 00:01:25,750 And we're going to know if it's a face while everything is not that simple. 19 00:01:25,760 --> 00:01:38,930 The reason for this hurdle is that even in a 24 by 24 pixel image the number of features is huge the 20 00:01:38,940 --> 00:01:45,810 number of features actually is over a hundred and eighty thousand possible features that can fit into 21 00:01:45,810 --> 00:01:46,340 this image. 22 00:01:46,350 --> 00:01:52,650 Sounds it sounds unbelievable because the image is so small it's only 24 pixels but it is true. 23 00:01:52,650 --> 00:01:58,800 So the the reason for that is that these features as we discussed there are scalable. 24 00:01:58,800 --> 00:02:03,810 So not only are we looking at these features on all their different positions in this image. 25 00:02:03,810 --> 00:02:08,010 We're actually looking at all of the variations of each one of these features. 26 00:02:08,010 --> 00:02:13,380 For instance this feature looks like for example it looks like there's one 2 pixels here and 1 2 pixels 27 00:02:13,380 --> 00:02:13,840 here. 28 00:02:13,890 --> 00:02:19,860 So you need to look through this image and try all all possible positions of this feature or very likely 29 00:02:19,850 --> 00:02:24,880 like that and then all possible positions here here all different positions here. 30 00:02:25,230 --> 00:02:26,660 And that's that's the start. 31 00:02:26,700 --> 00:02:33,540 But now you need to also extend that as to be like one to three pixels three pixels here and three pixels 32 00:02:33,560 --> 00:02:38,380 here and again you have to try out all possible positions for that feature. 33 00:02:38,400 --> 00:02:41,730 Next you might make it like four pixels high. 34 00:02:41,800 --> 00:02:46,170 Again you have to try it out all the positions the next you might make it for pixels high and then two 35 00:02:46,170 --> 00:02:48,360 pixels wide so that one. 36 00:02:48,360 --> 00:02:52,920 One two three four one two three four here and one two three four one two three four here and you need 37 00:02:52,920 --> 00:02:54,640 to try that out as well. 38 00:02:54,780 --> 00:03:00,030 And then once and then you keep doing that you keep expanding making it wider and so on and all the 39 00:03:00,030 --> 00:03:03,260 possible widths and heights of this feature. 40 00:03:03,450 --> 00:03:07,920 And then once you've run out of options for that you need to move on to the next one and the next or 41 00:03:07,920 --> 00:03:08,960 next on THE NEXT ONE. 42 00:03:09,150 --> 00:03:17,740 And so in a 24 by 24 pixel image these base features they all possible variations of them and their 43 00:03:17,760 --> 00:03:18,650 different positions. 44 00:03:18,810 --> 00:03:26,670 They add up to over a hundred and eighty thousand possible options that we would have to explore. 45 00:03:26,880 --> 00:03:29,670 And that is a huge number. 46 00:03:29,710 --> 00:03:33,210 So and it poses two concerns. 47 00:03:33,210 --> 00:03:40,550 First of all during the training that would be very hard because you know not only have to check 180000 48 00:03:40,560 --> 00:03:47,520 features for one image you need to check hundred eighty thousand features for all the images in the 49 00:03:47,520 --> 00:03:53,670 training data which is nine thousand eight hundred thirty two faces in the original Old Jones paper 50 00:03:54,060 --> 00:04:02,730 plus the huge number you know thousands and tens of thousands of images of known faces so you'd have 51 00:04:02,730 --> 00:04:10,830 to check that across all those images so training becomes quite long becomes like a nightmare. 52 00:04:10,950 --> 00:04:18,900 And the second thing is that even during your application so when you're detecting you're in if you've 53 00:04:18,900 --> 00:04:24,990 somehow managed to train those all those features now when you're detecting the faces you have to check 54 00:04:25,000 --> 00:04:27,260 180000 every single time. 55 00:04:27,480 --> 00:04:29,540 And that is practically impossible to do in real time. 56 00:04:29,550 --> 00:04:37,670 Take a lot of computation to do that for every single frame or every single image. 57 00:04:37,850 --> 00:04:41,960 So this is where adaptive boosting comes in to help solve this problem. 58 00:04:41,960 --> 00:04:48,560 So we've got we're going to take our features and we're going to put them together into a classifier 59 00:04:48,830 --> 00:04:58,280 which will look like this so here on the left we've got the classifier f f of x and then here f one 60 00:04:58,280 --> 00:05:04,460 of two and three are the features and Alpha 1 2 and 3 are the way to those features so we'll just add 61 00:05:04,460 --> 00:05:05,750 some features here. 62 00:05:05,750 --> 00:05:14,330 For instance there's a feature so that feature could be that's one that's commonly talked about the 63 00:05:16,010 --> 00:05:19,840 bridge of the nose of the nose is usually lighter than on the left on the right. 64 00:05:19,840 --> 00:05:29,430 So it's it's a feature that can help detect faces and this is the feature that the eyes are commonly 65 00:05:29,430 --> 00:05:32,230 darker than the area under the eyes. 66 00:05:32,310 --> 00:05:35,970 And so just even these two features will get as I say. 67 00:05:35,980 --> 00:05:40,110 So that's those two features and then maybe there's another feature and so on. 68 00:05:40,110 --> 00:05:46,170 So we've got our features aligned and that and it keeps going like that. 69 00:05:46,410 --> 00:05:51,400 And each one of these features is called a weak classifier on its own. 70 00:05:51,450 --> 00:05:57,840 He doesn't get a very high rate of success so as long as it gets over 50 percent that's already good. 71 00:05:57,840 --> 00:06:03,750 So for instance maybe this bridge of the nose if you just used that feature on its own maybe gets a 72 00:06:03,750 --> 00:06:10,560 60 percent success rate or a road or 65 percent success rate out of the images. 73 00:06:10,830 --> 00:06:19,560 And then this image that or this classifier or this part on the left the effort the F couple of X that's 74 00:06:19,560 --> 00:06:21,100 called a strong cost classified. 75 00:06:21,390 --> 00:06:28,160 And the way it works is that when you have one classifier by itself it's not as good. 76 00:06:28,170 --> 00:06:31,470 It maybe has a 60 percent success rate. 77 00:06:31,740 --> 00:06:38,430 When you have two week classifiers together even though this one might also be only like 60 or something 78 00:06:38,430 --> 00:06:39,970 percent or 55 percent. 79 00:06:40,050 --> 00:06:42,360 But together all of a sudden they're much stronger. 80 00:06:42,390 --> 00:06:48,730 And then as you keep adding they get stronger and stronger and stronger so you don't actually need. 81 00:06:48,990 --> 00:06:51,890 You don't need all hundred and eighty thousand of them. 82 00:06:52,080 --> 00:06:59,250 You might just need a couple of thousand to get a really really good result very strong classified. 83 00:06:59,400 --> 00:07:05,670 And this is called an ensemble method this is called ensemble because you are leveraging the power of 84 00:07:05,670 --> 00:07:06,930 the crowd as they call it. 85 00:07:06,930 --> 00:07:13,030 So it's like the power of one is not as strong. 86 00:07:13,320 --> 00:07:18,960 But when you put lots of we classifiers together they become a strong classifier and that is what I 87 00:07:18,960 --> 00:07:24,980 was going to say that even just these two features the bridge of the nose and the eyes and the area 88 00:07:24,990 --> 00:07:26,290 under the eyes. 89 00:07:26,730 --> 00:07:32,550 It's the exact numbers in the village on paper but I think they said something like it gave them an 90 00:07:32,610 --> 00:07:39,130 80 percent accuracy just those two features together or really made it so much more powerful. 91 00:07:39,180 --> 00:07:45,420 Not it wasn't close to 90 or a hundred but it was already much better than each one of those on their 92 00:07:45,420 --> 00:07:47,000 own. 93 00:07:47,160 --> 00:07:51,700 And so as you can imagine by adding more and more and more you can get better and better better results. 94 00:07:51,810 --> 00:07:57,840 So the question is now how do we find these right features how do we add the right ones. 95 00:07:57,870 --> 00:08:03,060 So the most important ones are want to have the most important ones at the front. 96 00:08:03,570 --> 00:08:04,050 And then. 97 00:08:04,080 --> 00:08:11,070 So then because our of the 180 thousand then the ones at the end you don't even need to worry about 98 00:08:11,070 --> 00:08:12,500 them after a certain point. 99 00:08:12,750 --> 00:08:16,680 So how do we find the best ones and this is where adaptive boosting comes into play. 100 00:08:16,680 --> 00:08:20,580 We won't go into the mathematics behind the algorithm we'll just go into the intuition. 101 00:08:20,850 --> 00:08:26,300 So let's say you have 10 pictures you have five faces and five known faces. 102 00:08:26,550 --> 00:08:31,520 And during the training process you apply. 103 00:08:31,540 --> 00:08:35,700 You want to identify the like the best feature. 104 00:08:35,700 --> 00:08:41,310 See I don't you want to identify how to build that strong classifier So you identify a feature that 105 00:08:41,310 --> 00:08:42,660 is important. 106 00:08:42,840 --> 00:08:52,060 And for instance the bridge of the nose feature and there you you apply it to your images. 107 00:08:52,290 --> 00:08:58,290 And so you get a result for for instance it says identify that these three are indeed faces. 108 00:08:58,290 --> 00:08:59,170 That's great. 109 00:08:59,250 --> 00:09:02,220 And then it's identified out of the ones on the right. 110 00:09:02,220 --> 00:09:05,020 It's identified that these three are not. 111 00:09:05,160 --> 00:09:08,520 And hasn't found that feature in those images. 112 00:09:08,520 --> 00:09:14,700 So for the for the algorithm those are not faces that's also good but then it does have some error. 113 00:09:14,700 --> 00:09:20,100 So it's got some these false negatives. 114 00:09:20,100 --> 00:09:24,170 Right so it's it's a Negat saying it's a negative but it's actually false. 115 00:09:24,210 --> 00:09:26,170 It's not so correct. 116 00:09:26,530 --> 00:09:31,620 These got two false negatives that says these are not faces when they are actually faces so didn't find 117 00:09:31,620 --> 00:09:34,650 the feature on these on these two pictures. 118 00:09:34,650 --> 00:09:36,360 But these are actually faces. 119 00:09:36,960 --> 00:09:39,450 And then it's got two false positives. 120 00:09:39,450 --> 00:09:46,760 It's identified these as faces even though they're not faces so what adaptive boosting does is it says 121 00:09:46,790 --> 00:09:47,390 ok cool. 122 00:09:47,390 --> 00:09:56,150 So we we like this feature because and the way we found it was perhaps that it is present on a huge 123 00:09:56,150 --> 00:10:04,850 number of images it's the one that comes up most commonly and through our like empirical training we've 124 00:10:04,850 --> 00:10:10,120 identified that yes this is a good picture but now what adaptive boosting will do is. 125 00:10:10,150 --> 00:10:16,070 OK I need to compliment like I might have a lot of very good features I might have this feature is good 126 00:10:16,170 --> 00:10:18,050 then another feature is good and so on. 127 00:10:18,050 --> 00:10:24,350 So but our idea is to go back here so the point of adaptive boosting is that we don't want to just combine 128 00:10:24,350 --> 00:10:31,220 the strongest features or because they might they might actually be leveraging off very similar things 129 00:10:31,220 --> 00:10:36,050 or that might not be the best approach the best approach that we want to take this is the important 130 00:10:36,050 --> 00:10:41,360 part is we want to take a strong feature but then we want to complement it with something that will 131 00:10:41,360 --> 00:10:46,140 help you know help fix where this feature is strong. 132 00:10:46,160 --> 00:10:49,320 That's good but where it's weak where it's making mistakes. 133 00:10:49,430 --> 00:10:57,800 This will help complement that area and will help improve the performance in those areas which are falling 134 00:10:57,800 --> 00:11:00,060 behind in this feature by itself. 135 00:11:00,230 --> 00:11:04,580 And then this feature will be used to complement these two. 136 00:11:04,700 --> 00:11:06,180 And we're they're falling behind. 137 00:11:06,170 --> 00:11:11,060 And so each next one we're not just taking the strongest features we can find we're staying a strong 138 00:11:11,060 --> 00:11:16,180 feature or the strongest and then we take in the one that best complements this one and then we're taking 139 00:11:16,220 --> 00:11:19,610 the next one which best complements these two and the next on and so on. 140 00:11:19,610 --> 00:11:26,150 So basically instead of just taking the strongest all the time we're constructing the strongest classifier 141 00:11:26,300 --> 00:11:30,620 to resulting classify that we can rather than. 142 00:11:30,620 --> 00:11:38,390 And so basically by leveraging the strengths of this one and then fixing up its weaknesses with this 143 00:11:38,390 --> 00:11:42,380 one and then fixing up their weaknesses with this and so on so we covered. 144 00:11:42,400 --> 00:11:45,420 That's very kind of like a strategic approach. 145 00:11:45,500 --> 00:11:53,630 And so here again we've got these are this is all good but then we've got these two false negative false 146 00:11:53,900 --> 00:11:56,420 negatives and false positives. 147 00:11:56,480 --> 00:12:04,790 And so what adaptive boosting our will do is in the next round it will now look for something that complements 148 00:12:04,880 --> 00:12:08,190 this feature that we found and how we'll do it. 149 00:12:08,200 --> 00:12:12,890 What I'll say is it will give more weight to where the errors were made. 150 00:12:12,890 --> 00:12:16,630 Ill say Okay so now I'm going to let's go let's go back. 151 00:12:16,640 --> 00:12:21,570 Although explain this or so it's decreased the sizes. 152 00:12:21,620 --> 00:12:28,010 This is just to just to symbolize what is going on so we've decreased the size of these images increase 153 00:12:28,010 --> 00:12:32,630 the sizes of these zones highlighted in blue and decreased the size of these on increase the size of 154 00:12:32,630 --> 00:12:36,170 these ones so you can see again if I go back. 155 00:12:36,170 --> 00:12:37,290 So there you go. 156 00:12:37,280 --> 00:12:44,180 So it's increased the sizes of these ones saying that it's not actually increasing the size of the image 157 00:12:44,180 --> 00:12:47,790 which is doing it here on the PowerPoint just so that it is clear. 158 00:12:47,810 --> 00:12:52,790 It's like it's easier to follow along but what it's doing is just increasing the weight the importance 159 00:12:52,790 --> 00:13:01,430 of these images for the next for the next horror like feature one is going to be looking at this whole 160 00:13:01,730 --> 00:13:07,970 system is going to now is going to know that the importance of these images is higher and it's going 161 00:13:07,970 --> 00:13:14,420 to find a feature that treats them the best that helps I classify them properly. 162 00:13:14,660 --> 00:13:15,850 And that's what we do. 163 00:13:15,860 --> 00:13:24,770 That's how it accounts for that is how it implements that notion where we want to fix the problems of 164 00:13:24,770 --> 00:13:27,720 the first class fire the first feature that we had so. 165 00:13:27,950 --> 00:13:35,030 So now if we add a new one it finds the next one that best works with these images that predominantly 166 00:13:35,030 --> 00:13:37,550 classify these images properly and find this one. 167 00:13:37,550 --> 00:13:42,920 And so as you can see this now it applies to classify these five images correctly. 168 00:13:43,100 --> 00:13:46,950 Unfortunately classify that or an incorrectly so that's a false negative. 169 00:13:47,300 --> 00:13:50,470 And then it's classified these images correctly and then it's classified. 170 00:13:50,480 --> 00:13:53,260 This one still incorrectly it's a false positive. 171 00:13:53,270 --> 00:14:01,850 So maybe it was not possible to find a classifier that would classify these two very well so what the 172 00:14:01,850 --> 00:14:05,960 algorithm did is it found one that classifies these four including this one very well. 173 00:14:06,140 --> 00:14:11,670 And now we're going to have to fix find a new classifier that fixes these errors. 174 00:14:11,870 --> 00:14:13,250 And that's what will happen next. 175 00:14:13,250 --> 00:14:16,240 Next we would focus on these two. 176 00:14:16,490 --> 00:14:21,380 So everything will become smaller and would focus on these two as predominantly the ones that we need 177 00:14:21,380 --> 00:14:28,070 to fix in order to create that very strong overall classifier. 178 00:14:28,070 --> 00:14:29,750 And and this would keep going. 179 00:14:30,080 --> 00:14:37,070 And in the end the goal is to create a classifier that classifies all of these correctly and classifies 180 00:14:37,100 --> 00:14:39,230 all of these correctly as well. 181 00:14:39,330 --> 00:14:41,440 These faces spaces these are known faces. 182 00:14:41,480 --> 00:14:44,110 So of course it's not never going to be ideal. 183 00:14:44,120 --> 00:14:53,150 But the point is to keep building this strong classifier through these weak classifiers just keep building 184 00:14:53,150 --> 00:15:01,570 it and minimizing the error rate until you get to a satisfactory level until you get to a very high 185 00:15:03,550 --> 00:15:08,890 percentage of correct classifications like 99 point something percent. 186 00:15:08,890 --> 00:15:14,980 And at that point you don't need the hundred and eighty thousand You just need that first couple of 187 00:15:14,980 --> 00:15:20,530 thousand that or hundreds however many you can get to. 188 00:15:20,550 --> 00:15:27,400 You mean you'll need to get to that good satisfactory conversion rate and then that's it. 189 00:15:27,400 --> 00:15:29,830 And that's all you need. 190 00:15:30,060 --> 00:15:31,870 And then you don't classify as really. 191 00:15:32,020 --> 00:15:43,720 And that's a first step to reducing the number of the computation expense of finding find these images. 192 00:15:43,720 --> 00:15:48,420 That's the first step is the added boost in what we discussed today. 193 00:15:48,610 --> 00:15:52,010 And then the next step is going to be the cascading. 194 00:15:52,240 --> 00:15:58,050 And once you combine them together you'll see that it actually improves the efficiency substantially. 195 00:15:58,360 --> 00:16:02,380 So there you go that's how the Arab Boosler adaptive boosting algorithm works. 196 00:16:02,530 --> 00:16:11,120 If you'd like to read a bit more about it there is a paper by Gene you and Paul Viola. 197 00:16:11,140 --> 00:16:12,910 It's called boosting image retrieval. 198 00:16:12,910 --> 00:16:19,690 And this paper inspired some of the work in the violin Jones paper which we were talking about at the 199 00:16:19,690 --> 00:16:20,370 start. 200 00:16:20,560 --> 00:16:23,220 So hope you enjoyed Sterle 40 next 14. 201 00:16:23,380 --> 00:16:25,540 And until then enjoy computer vision. 22897