1 00:00:00,330 --> 00:00:00,990 Welcome back. 2 00:00:01,230 --> 00:00:03,480 This is a lecture on YOLO. 3 00:00:03,540 --> 00:00:09,870 This is an overview, a high level overview of YOLO, which is my favorite object detector, so let's 4 00:00:09,870 --> 00:00:10,590 get started. 5 00:00:11,880 --> 00:00:17,730 So your stands for you only look once and was developed by two brilliant researchers, Joseph Redmond 6 00:00:17,730 --> 00:00:20,190 and Ali Farad in 2015. 7 00:00:20,310 --> 00:00:26,250 And this is basically a diagram that has become famous for YOLO because its use illustrates how good 8 00:00:26,250 --> 00:00:27,870 and how simple it was. 9 00:00:28,890 --> 00:00:34,890 So YOLO as a single stage detector, hence the name you only look once single stage, meaning that it 10 00:00:34,890 --> 00:00:40,010 actually saves a lot of time compared to our CNN's and even faster CNN's. 11 00:00:40,440 --> 00:00:45,750 Because the biggest drawback with faster our CNN's was that it was quite slow. 12 00:00:45,870 --> 00:00:53,940 It wasn't practical to run this on my video because frame rate on even powerful GPUs use back then with 13 00:00:53,940 --> 00:00:59,430 single digit frame rates, it was a seven usually unloved or faster and video super GPUs at that time. 14 00:01:00,330 --> 00:01:06,690 So since these two did dramatically improve that and they were also using a single stage technique, 15 00:01:06,690 --> 00:01:10,530 however, they wouldn't as nearly as accurate as our CNN's. 16 00:01:11,100 --> 00:01:13,350 And that's where YOLO is so important. 17 00:01:13,890 --> 00:01:14,430 YOLO. 18 00:01:14,550 --> 00:01:19,200 And now you look for and you live five attempted to solve both of those problems and strike a balance 19 00:01:19,200 --> 00:01:20,760 between speed and accuracy. 20 00:01:21,210 --> 00:01:28,200 And I actually would say no YOLO which and four and five and your X surpasses all other option factors 21 00:01:28,650 --> 00:01:33,150 in not just performance, but I mean, not not at speed, but in accuracy as well. 22 00:01:34,140 --> 00:01:36,030 So there are many flavors of YOLO. 23 00:01:36,600 --> 00:01:41,700 YOLO has had tremendous success in real world applications and have many, many different visions right 24 00:01:41,700 --> 00:01:41,970 now. 25 00:01:42,690 --> 00:01:49,410 Some examples are Tinio your motivation to live within Shreeve, which is four or five you'll scale 26 00:01:49,410 --> 00:01:55,830 the YOLO YOLO with various backhands is very customizable, and the research world and research community 27 00:01:55,830 --> 00:02:00,960 has really taken advantage of that and has explored so many different visions of YOLO. 28 00:02:01,560 --> 00:02:05,970 Right now, the most popular flavor fuel will use an industry at least right. 29 00:02:06,840 --> 00:02:10,110 Maybe before 2019 was Univision tree. 30 00:02:10,140 --> 00:02:13,980 However, no, it's the ultimate x implementation of YOLO five. 31 00:02:14,490 --> 00:02:19,960 YOLO five, which we'll talk about later on in a few lectures, is technically YOLO four. 32 00:02:19,980 --> 00:02:26,910 However, it's been YOLO for ported to PyTorch, with a number of optimizations that make training quite 33 00:02:26,910 --> 00:02:28,560 quick inference testing. 34 00:02:28,920 --> 00:02:34,110 It's really, really well done by automatics, and actually, I encourage you to use that can go to 35 00:02:34,110 --> 00:02:35,280 the GitHub and check it out. 36 00:02:35,400 --> 00:02:38,610 I'll actually be teaching a lecture on that later on in the course anyway. 37 00:02:39,360 --> 00:02:45,330 And because of how common YOLO is, is a number of pre-trained your models on the data said that are 38 00:02:45,330 --> 00:02:48,270 readily available out there that you can use in your code. 39 00:02:48,300 --> 00:02:54,750 Right now, it's quite easy to use, and these are the easy classes that it's been trained on on the 40 00:02:54,750 --> 00:02:55,560 cocoa dataset. 41 00:02:56,490 --> 00:02:59,850 So how does your little compare to other objects adapters? 42 00:03:00,420 --> 00:03:01,450 Well, let's take a look. 43 00:03:01,470 --> 00:03:05,820 So currently, the best performing YOLO models are YOLO four. 44 00:03:06,300 --> 00:03:09,130 Technically, you'll have four and five are the same, but your five is a little. 45 00:03:09,330 --> 00:03:11,870 That's a generally and you'll X. 46 00:03:11,910 --> 00:03:14,030 Then you can see this was a from the publication. 47 00:03:14,040 --> 00:03:18,450 This image here, the top right one from the YOLO four people. 48 00:03:18,900 --> 00:03:25,110 And you can see the speeds here, and the V 100 CPU was magnificent. 49 00:03:25,110 --> 00:03:29,160 It was like 110, and it was still getting very high at accuracies. 50 00:03:29,730 --> 00:03:35,910 You can see the efficient detect model with the backends we're getting better Macnab scores. 51 00:03:36,580 --> 00:03:36,960 However, 52 00:03:41,070 --> 00:03:42,270 however, they were quite slow. 53 00:03:43,230 --> 00:03:47,610 And then you can see in this diagram here this is the bottom left program. 54 00:03:48,120 --> 00:03:49,080 Yellow five. 55 00:03:49,080 --> 00:03:52,380 This is the diagram we used to compare to efficient with text. 56 00:03:52,860 --> 00:03:58,530 You can see it's actually beating a lot of the efficient tech models in terms of faster inference and 57 00:03:58,530 --> 00:04:01,050 better maps go as well, if it's quite remarkable. 58 00:04:01,680 --> 00:04:07,980 And this diagram here was from the Yellow X research paper that again showed it speeding efficient detect 59 00:04:07,980 --> 00:04:11,670 light, though not the efficient detect, not non light visions. 60 00:04:12,750 --> 00:04:14,940 You can see this is a small vision here. 61 00:04:14,970 --> 00:04:20,750 This is a larger vision here as well, with using the full efficient detector model YOLO Vision five 62 00:04:20,750 --> 00:04:25,500 a different backhands as well, and large and small models. 63 00:04:26,070 --> 00:04:31,680 And you can see your Excel has achieved quite good map score as well. 64 00:04:32,790 --> 00:04:34,590 So how does your work? 65 00:04:34,800 --> 00:04:43,230 Well, in prior object detection systems, they repurposed classify as a localized localizes to perform 66 00:04:43,230 --> 00:04:48,690 detection, and then they applied the model to an image at multiple locations and skills. 67 00:04:49,200 --> 00:04:54,120 So high scoring regions are considered image considered optic detections. 68 00:04:54,960 --> 00:04:59,760 So instead, yellower plays a single neural network. 69 00:04:59,900 --> 00:05:06,190 The full image, this network divides the image into regions and predicts bounding boxes and probabilities 70 00:05:06,190 --> 00:05:10,450 for each region probabilities of different classes to be specific. 71 00:05:10,990 --> 00:05:17,110 These bounding boxes are then weighted by the predicted probabilities, and this method has several 72 00:05:17,110 --> 00:05:22,000 advantages over classifier based systems because it looks at the whole image. 73 00:05:22,180 --> 00:05:29,170 Test time So its predictions are informed by the global context of the image, and also it makes predictions 74 00:05:29,170 --> 00:05:31,540 with a single network evaluation. 75 00:05:31,990 --> 00:05:36,460 Unlike RC events, which require 2000 for a single image. 76 00:05:37,150 --> 00:05:43,450 So this makes it extremely fast more than a thousand times faster than the original or CNN and 100x 77 00:05:43,450 --> 00:05:48,820 faster and faster as the events and that coming to the end of this lesson. 78 00:05:48,850 --> 00:05:53,820 However, I really do encourage you to read the your vision three people. 79 00:05:53,830 --> 00:05:55,390 It's actually quite hilarious. 80 00:05:55,840 --> 00:06:01,300 The researchers definitely don't take themselves too seriously, and it's quite an entertaining read 81 00:06:01,420 --> 00:06:02,530 to be thought to be us. 82 00:06:03,130 --> 00:06:04,330 So we'll stop there. 83 00:06:04,480 --> 00:06:09,880 And in the next section, we'll take a deeper look on how exactly does your work? 84 00:06:10,390 --> 00:06:12,430 So I'll see you in the next section. 85 00:06:12,580 --> 00:06:13,000 Thank you.