1 00:00:00,180 --> 00:00:05,670 Welcome back in this section, we'll take a look at how we actually Trina YOLO model. 2 00:00:06,240 --> 00:00:07,350 So let's get started. 3 00:00:07,500 --> 00:00:14,130 So firstly, in all training processes, you have the training data, which includes the ground truth 4 00:00:14,130 --> 00:00:16,330 labels as well as a test data. 5 00:00:16,350 --> 00:00:20,090 But for now, we'll be focusing on how we use the training data in your. 6 00:00:20,640 --> 00:00:24,990 So initially, we have a human who annotate these these images here. 7 00:00:25,410 --> 00:00:29,640 So we have a ground truth green box here with the class being dug. 8 00:00:30,240 --> 00:00:37,440 So during training, we have to be basically tried to get the model to basically learn a function that 9 00:00:37,440 --> 00:00:41,940 can propose a box that is as close to this box as possible with the precise class. 10 00:00:42,930 --> 00:00:45,180 So let's take a look what happens during training? 11 00:00:45,720 --> 00:00:50,160 So during training, the model attempts to match the example to the right. 12 00:00:50,160 --> 00:00:50,550 So. 13 00:00:50,760 --> 00:00:57,450 So initially, we have the mapping being here and we have the plus class probability being like this 14 00:00:57,450 --> 00:00:57,670 here. 15 00:00:57,690 --> 00:01:00,180 So one being taught and the other classes being zero. 16 00:01:01,290 --> 00:01:05,070 Next, we get two bounding box predictions for that. 17 00:01:05,070 --> 00:01:05,430 So. 18 00:01:06,120 --> 00:01:09,480 So we get these two boxes here, as you can see in this example. 19 00:01:10,080 --> 00:01:14,130 However, we need to adjust these bounding boxes because they both can't be right. 20 00:01:14,670 --> 00:01:19,410 So you can see that is one the larger box is closer to the ground truth box here. 21 00:01:19,860 --> 00:01:25,500 So what we do, we increase the confidence of that box and then decrease the confidence of this box 22 00:01:25,500 --> 00:01:26,430 of smaller box. 23 00:01:26,850 --> 00:01:30,180 So by doing that, you can see we just made this larger here. 24 00:01:30,450 --> 00:01:33,480 Sorry, we made the box larger here on this one. 25 00:01:34,110 --> 00:01:39,120 And then simultaneously, we make that the middle box smaller, so we lower the confidence. 26 00:01:39,120 --> 00:01:40,980 So this is what we get here. 27 00:01:41,670 --> 00:01:48,300 So no cells with no ground to it because you will have cells that predict that basically bounding boxes 28 00:01:48,300 --> 00:01:49,000 over nothing. 29 00:01:49,030 --> 00:01:54,930 The background we call it those basically, we decrease the confidence of those and we don't adjust 30 00:01:54,960 --> 00:01:57,690 class probabilities or coordinates of these boxes. 31 00:01:58,560 --> 00:02:02,790 So that's essentially how Eula tries to fit these boxes to the data. 32 00:02:03,750 --> 00:02:05,790 But how do we how does it actually work? 33 00:02:06,270 --> 00:02:09,450 Well, we need a combination of three lost functions to do this. 34 00:02:09,990 --> 00:02:11,930 Firstly, we need a classification loss. 35 00:02:11,940 --> 00:02:18,150 So if an object is detected, it is a squared error loss of the class conditional probabilities for 36 00:02:18,150 --> 00:02:19,050 each class. 37 00:02:19,710 --> 00:02:25,350 Then we have the localization of loss, which measures the performance for the predicted both bounding 38 00:02:25,350 --> 00:02:27,270 box to the ground it. 39 00:02:27,900 --> 00:02:29,850 And then we have the confidence loss. 40 00:02:30,240 --> 00:02:33,150 That's the confidence that the box has an object. 41 00:02:33,750 --> 00:02:40,010 So that's basically a short summary of how the training process in Yuma works, and you can see just 42 00:02:40,020 --> 00:02:42,000 a summary slide after you lose performance. 43 00:02:42,000 --> 00:02:44,190 This is the a little bit and treat, to be precise. 44 00:02:44,760 --> 00:02:51,210 You can see how much faster it was 45 frames per second compared to the same 50 yards at the time, 45 00:02:51,210 --> 00:02:56,790 which was faster RC and just getting a better, much better map score by at least 10 points. 46 00:02:57,180 --> 00:03:04,140 However, it was a lot slower than usual, and you can see you're going does generalize and generalize 47 00:03:04,140 --> 00:03:10,740 quite well, even on other paintings like this than defies bottles dining table Pearson's cat? 48 00:03:11,310 --> 00:03:12,150 So that's quite cool. 49 00:03:12,150 --> 00:03:15,360 And by the way, the original Mona Lisa painting did not have a cut. 50 00:03:15,450 --> 00:03:21,690 In case you were wondering also, you know, compared to past our CNN's, you can see that actually 51 00:03:21,690 --> 00:03:26,340 fast approaching ends, at least compared to your vision tree did perform better. 52 00:03:26,490 --> 00:03:29,400 You can see 71 percent correct as opposed to 65. 53 00:03:30,000 --> 00:03:31,920 However, Eula was much faster. 54 00:03:32,010 --> 00:03:39,150 So the key takeaways from this your lesson is that you lose fast and you lose vision for Vision five 55 00:03:39,150 --> 00:03:45,330 and except perhaps the best in accuracy right now in 2001, still surpassing other models which include 56 00:03:45,660 --> 00:03:48,960 our CNN's detector on to which we'll talk about shortly. 57 00:03:49,320 --> 00:03:54,330 And if we shouldn't detect the other provides end to end training, which we've seen there, and it 58 00:03:54,330 --> 00:03:59,940 gives us very little background error and even doing the illusion treat him up wasn't as good as our 59 00:03:59,940 --> 00:04:00,540 CNN's. 60 00:04:00,900 --> 00:04:02,130 It was a lot faster. 61 00:04:02,550 --> 00:04:06,240 When you look does tend to have more localization errors at times. 62 00:04:06,240 --> 00:04:11,430 However, in the little regions of your world, which is you mentioned four or five in X, they actually 63 00:04:11,430 --> 00:04:15,660 have minimized that quite a bit, so it will stop them from now. 64 00:04:16,170 --> 00:04:23,490 And in the next section, we'll take a look at the architecture and evolution from region three to version 65 00:04:23,490 --> 00:04:23,880 five. 66 00:04:24,570 --> 00:04:25,620 So stay tuned for that. 67 00:04:25,740 --> 00:04:26,520 Thank you for watching.