Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,690 --> 00:00:04,830 Hello and welcome back to the course on computer vision got an exciting section ahead and we're going 2 00:00:04,830 --> 00:00:13,470 to kick off by talking about how SS D or the single shot multi box detection algorithm is different 3 00:00:13,550 --> 00:00:15,260 to others. 4 00:00:15,440 --> 00:00:16,840 This looking at this image. 5 00:00:16,860 --> 00:00:21,860 So here is an image and. 6 00:00:22,130 --> 00:00:22,550 Yeah. 7 00:00:22,680 --> 00:00:26,060 So we can see some sheep walking around on a field. 8 00:00:26,250 --> 00:00:33,500 And the question is how would normally a computer or object detection algorithms detect these sheep 9 00:00:34,520 --> 00:00:42,190 Well it's the sheep as you can see quite small and you not only need to look at the image as we learned 10 00:00:42,200 --> 00:00:46,580 in the Anik some computer unconventional neural networks there where we're just looking at image and 11 00:00:46,580 --> 00:00:51,170 saying if something is present on the image or not here we actually need to detect exactly where the 12 00:00:51,170 --> 00:00:56,840 sheep are and put those boxes around them to you know say that the sheep is here or the sheep is here 13 00:00:56,840 --> 00:01:05,800 and so on and so what normally these algorithms would do is they would try to save time or try to save 14 00:01:05,810 --> 00:01:14,900 the computational cost by through certain tricks and hacks and of course all of them even less as D 15 00:01:14,900 --> 00:01:16,120 has its own trick and hacks. 16 00:01:16,130 --> 00:01:23,480 But what changes is the tricks and hacks that those things change and so previously most are all like 17 00:01:23,480 --> 00:01:25,490 a huge portion of the algorithms. 18 00:01:25,490 --> 00:01:31,870 What they would do is they would engage in something called an object proposal methodology. 19 00:01:31,880 --> 00:01:36,890 They would have some object proposal techniques and basically you've got an image and of course is a 20 00:01:36,890 --> 00:01:44,150 huge image Normally these images are resized to smaller images and they were like the algorithms work 21 00:01:44,150 --> 00:01:47,050 with smaller images and then maybe they resized back in and so on. 22 00:01:47,060 --> 00:01:53,310 But it is like even on a small image you still have the same problem that if you were to for instance 23 00:01:53,330 --> 00:01:58,520 convert this whole image with rectangle and then try to you know try to guess in which one there's a 24 00:01:59,000 --> 00:02:03,440 sheep or not let's say here's a rectangle Here's a rectangle you can see that most of them don't actually 25 00:02:03,440 --> 00:02:10,790 have the sheep and you would be spending a lot of time finding these these sheep and a lot of can be 26 00:02:10,790 --> 00:02:15,490 additional cost because of competition cost and Dalgard would be slow and that's a big problem. 27 00:02:15,740 --> 00:02:21,830 And so what they would normally do is do this thing called object proposal where they would come up 28 00:02:21,830 --> 00:02:30,290 with a way to break down the image segmented into parts to suggest where they could potentially be objects 29 00:02:30,290 --> 00:02:35,720 and where they could not be objects for instance by looking at the gradients in color. 30 00:02:35,840 --> 00:02:39,800 So if you look at this big chunk over here it's all green. 31 00:02:39,800 --> 00:02:47,510 There's not much there's not much gradient there's not much changes not much chare like it's all gradual 32 00:02:47,510 --> 00:02:51,950 and there's no red there's no edges there's no there's a little bit of edges but they're all kind of 33 00:02:51,950 --> 00:02:56,990 like it's all about the same kind of texture same same color more or less. 34 00:02:57,230 --> 00:02:58,390 And so there's no. 35 00:02:58,580 --> 00:03:04,700 Nothing changes radically even from like one neighboring pixel to the next and therefore you could make 36 00:03:04,700 --> 00:03:07,560 a proposal that there is no object in here. 37 00:03:07,730 --> 00:03:12,770 And then you would take this and you know maybe you would make a proposal that like it depends on how 38 00:03:12,770 --> 00:03:17,270 you break it up break it up like I can this you might say you know there's an object here then you might 39 00:03:17,270 --> 00:03:20,220 say that there's an object here because you know you can see the shadow. 40 00:03:20,330 --> 00:03:26,870 And again we're not just detecting sheep we're detecting idealy in the most complex of all challenges 41 00:03:26,870 --> 00:03:32,420 predicting any types of objects that were officers so for instance you know sheep or we could be looking 42 00:03:32,420 --> 00:03:38,900 for mice and we could be looking for cars and we could look looking for giraffes anything. 43 00:03:38,900 --> 00:03:44,300 So it's very dangerous to discard anything any object because then you might miss out. 44 00:03:44,460 --> 00:03:50,480 And then on the other hand if you take this segment off if it's segmented like that or however it's 45 00:03:50,480 --> 00:03:54,950 segment you might you would see that you would first calculate the gradient that there is some kind 46 00:03:54,950 --> 00:04:00,380 of change happening over here and there is an object proposal for this segment and then we will dig 47 00:04:00,380 --> 00:04:01,530 into this further. 48 00:04:02,000 --> 00:04:06,420 And so that's that way they save time. 49 00:04:06,480 --> 00:04:15,050 But as you can imagine it's they would sacrifice accuracy because there's all lots of other details 50 00:04:15,050 --> 00:04:21,470 to it but in short they would say can be additional cost but they would sacrifice accuracy and you know 51 00:04:21,470 --> 00:04:27,620 Miss objects sometimes and when it's kind of like it might be OK in some implication in other applications 52 00:04:27,630 --> 00:04:32,480 the like very very high percentage of accuracy especially in self-driving cars and things like that 53 00:04:32,480 --> 00:04:37,970 you can't afford to sacrifice accuracy but at the same time you also need for computational efficiency. 54 00:04:38,030 --> 00:04:42,740 So because it needs to happen all in all these happen real time. 55 00:04:42,770 --> 00:04:45,150 Yeah and so what did S's come up with. 56 00:04:45,170 --> 00:04:52,070 Well let's have a look at this slide is the they are the authors of SSD. 57 00:04:52,070 --> 00:04:57,150 They came up with a brand new solution. 58 00:04:57,260 --> 00:05:02,390 Well it's a very interesting solution where they do everything in one shot. 59 00:05:02,390 --> 00:05:04,540 That's why it's called the single shot. 60 00:05:04,660 --> 00:05:06,610 Box detection algorithm. 61 00:05:06,630 --> 00:05:08,100 It it can. 62 00:05:08,150 --> 00:05:11,420 It just looks at the image once it doesn't have to go back to it. 63 00:05:11,410 --> 00:05:17,800 It doesn't have to do this object proposal it doesn't have to then run many convolutional neural networks. 64 00:05:17,810 --> 00:05:20,030 That's the part where the communicational cost comes in. 65 00:05:20,030 --> 00:05:28,850 So if you go back one slide here if you like break it down and then you run a of neural network for 66 00:05:28,850 --> 00:05:30,920 every single one of those rectangles to detect. 67 00:05:30,920 --> 00:05:35,220 OK is there a ship there is there a ship there is there a ship there is this ship that that's very can 68 00:05:35,220 --> 00:05:40,350 be really costly and that's where you need to do these jobs proposals to cut out parts that are definitely 69 00:05:40,440 --> 00:05:45,430 out of the question so you don't have to run Minicom the conventional network many times. 70 00:05:45,510 --> 00:05:52,220 But in the single shot multi-book detection algorithm what they did is they only look at it once there's 71 00:05:52,230 --> 00:05:55,920 only one algorithm one other algorithm in the world that does the same thing. 72 00:05:56,010 --> 00:05:58,130 It's called the YOLO algorithm. 73 00:05:58,710 --> 00:06:05,310 It's stands for you only look once and in their paper they actually compare the two. 74 00:06:05,390 --> 00:06:07,110 This is does a better job. 75 00:06:07,110 --> 00:06:10,080 They prove empirically that Issas is better. 76 00:06:10,140 --> 00:06:15,810 But apart from that apart from yellow and the there is no algorithm other algorithm in the world that 77 00:06:16,440 --> 00:06:23,520 or at the time there was no other algorithm in the world that actually only looks at the image ones 78 00:06:23,550 --> 00:06:24,620 and that that's it. 79 00:06:24,610 --> 00:06:29,100 It only does that one can volitional neural networks all of them would come back to the image several 80 00:06:29,100 --> 00:06:31,410 times or many times. 81 00:06:31,560 --> 00:06:33,870 And one second. 82 00:06:34,100 --> 00:06:34,850 OK. 83 00:06:35,340 --> 00:06:36,420 And. 84 00:06:36,530 --> 00:06:36,840 Yeah. 85 00:06:36,840 --> 00:06:39,900 And so now we have this and that's we're going what did they do. 86 00:06:39,900 --> 00:06:46,450 So basically all of those boxes as you will we will discuss in the further trials will the boxes and 87 00:06:46,530 --> 00:06:51,120 boxes that are up here on the image all of them go through the network at the same time. 88 00:06:51,120 --> 00:06:57,950 And moreover not just to go through the network at the same time network also remembers which boxes 89 00:06:58,020 --> 00:06:59,560 dealing with the boxes. 90 00:06:59,640 --> 00:07:08,460 They train up on detecting objects and detecting the borders or objects correctly. 91 00:07:08,880 --> 00:07:16,040 And moreover in this same network and now the really cool thing that they introduced is that there's 92 00:07:16,050 --> 00:07:22,590 many convolutions to reduce the image size so you can see stars with 300 image 38 pixels 19 10 5 3 1 93 00:07:22,880 --> 00:07:24,460 as the image size of the juices. 94 00:07:24,600 --> 00:07:30,960 And that helps deal with different sizes of objects on the original image and we'll talk about that 95 00:07:30,960 --> 00:07:38,190 as well so they combined lots of hacks in this one algorithm and probably one of the most important 96 00:07:38,190 --> 00:07:41,690 ones is of course that they do everything in a single shot. 97 00:07:42,320 --> 00:07:49,500 All these it all happens in one huge convolution as it goes through this network and that really helps 98 00:07:49,500 --> 00:07:51,000 save on the communication efficiency. 99 00:07:51,210 --> 00:07:56,440 So that's in a nutshell what how this is different. 100 00:07:56,520 --> 00:07:59,790 So it does everything in a single shot. 101 00:07:59,790 --> 00:08:08,610 It can deal with training of identifying objects and also identifying the borders and it can deal with 102 00:08:08,960 --> 00:08:11,350 scale all all at the same time. 103 00:08:11,670 --> 00:08:20,850 And there's only one other algorithm that rival the D in that architecture and that set up it's called 104 00:08:20,850 --> 00:08:22,130 the yoa algorithm. 105 00:08:22,130 --> 00:08:27,300 And yeah and they talk about more in the paper so that's I'm not sure we'll go into those components 106 00:08:27,300 --> 00:08:31,860 in more detail in upcoming tutorials but we'll go into them in a very intuitive sense if you would like 107 00:08:31,860 --> 00:08:38,780 to get the math behind it all the more strict approach to what's going on. 108 00:08:38,790 --> 00:08:47,360 The best place to look at is of course the original paper by way you and others is called Single Shot 109 00:08:47,370 --> 00:08:51,870 multi box detector you can find our archive there's a link. 110 00:08:51,980 --> 00:08:57,480 And yeah this is quite a it's a very new algorithm and so therefore there isn't much written about it 111 00:08:58,360 --> 00:09:05,200 but nevertheless this paper is not as you can imagine the ultimate source of truth and into its Gregorie 112 00:09:05,850 --> 00:09:08,040 and on that note I hope you enjoyed this tutorial. 113 00:09:08,040 --> 00:09:12,420 Hope you're excited about the trials coming up in the section and I look for the next step. 114 00:09:12,420 --> 00:09:14,310 And until then enjoy computer vision. 13273