1 00:00:01,000 --> 00:00:01,000 Hi guys. 2 00:00:01,000 --> 00:00:03,000 Welcome to the new video tutorial. 3 00:00:03,000 --> 00:00:11,000 In this video we will discuss is YOLO V8, the real state of the art and will present a comparison between 4 00:00:11,000 --> 00:00:13,000 YOLO V8 and YOLO. 5 00:00:13,000 --> 00:00:18,000 V7 Before we move ahead, just a quick overview of YOLO. 6 00:00:18,000 --> 00:00:20,000 How does YOLO works? 7 00:00:20,000 --> 00:00:27,000 And so Joseph Redmon, Santos, Santosh Daruwala, Ross Girshick and Ali Farhadi introduced YOLO, which 8 00:00:27,000 --> 00:00:35,000 stands for You Only Look Once YOLO gained popularity because of its accuracy while maintaining a small 9 00:00:35,000 --> 00:00:42,000 model size from version 1 to 4, YOLO was maintained in a C code in a custom deep learning framework 10 00:00:42,000 --> 00:00:44,000 written by Redmond called Darknet. 11 00:00:44,000 --> 00:00:52,000 In the last two years, YOLO are YOLO, YOLO, v6, Yolo V7 have emerged around the world out of their 12 00:00:52,000 --> 00:00:55,000 own PyTorch based implementation. 13 00:00:55,000 --> 00:01:01,000 Each model has bought new SOTA techniques that combine to push model accuracy and efficiency. 14 00:01:01,000 --> 00:01:05,000 So let's move towards the objectives part. 15 00:01:08,000 --> 00:01:09,000 Upon reading our objectives. 16 00:01:09,000 --> 00:01:14,000 Or you can say that we have splitted the whole lecture into five major parts. 17 00:01:14,000 --> 00:01:21,000 So the first part is or the first objective is we will see how it is better than previous versions of 18 00:01:21,000 --> 00:01:22,000 YOLO. 19 00:01:22,000 --> 00:01:26,000 In the next part I will present a comparison of YOLO V8 with YOLO. 20 00:01:26,000 --> 00:01:30,000 V7 will see drink and license plate detection problem. 21 00:01:30,000 --> 00:01:38,000 We will do the training time comparison mean average procedure comparison of YOLO V8 with YOLO V7 In 22 00:01:38,000 --> 00:01:44,000 the third part I will do the result analysis which for the problem of license plate detection using 23 00:01:44,000 --> 00:01:52,000 YOLO V8 and YOLO V7 In the next part we will see what improvements are made in the YOLO V8 version than 24 00:01:52,000 --> 00:01:53,000 the previous versions. 25 00:01:53,000 --> 00:01:57,000 And my observations are in the last regarding the YOLO V8 version. 26 00:01:57,000 --> 00:02:03,000 So let's move towards the how YOLO V8 is better than previous versions of YOLO. 27 00:02:05,000 --> 00:02:06,000 Well, you will be. 28 00:02:06,000 --> 00:02:13,000 It introduces a new backbone network, which is Darknet 53, which is significantly faster and more 29 00:02:13,000 --> 00:02:19,000 accurate than the previous backbone used in Yolov7 or in the others versions. 30 00:02:19,000 --> 00:02:26,000 Darknet, 53, is a convolutional neural network that is 53 layers deep and can classify image into 31 00:02:26,000 --> 00:02:33,000 up to 1000 object categories such as keyboard, mouse, pencil and many other categories as well. 32 00:02:33,000 --> 00:02:40,000 Yolo V makes bounding box prediction similar to image segmentation which which is basically pixel wise. 33 00:02:41,000 --> 00:02:46,000 To achieve this they have introduce anchor free detection had the concept of YOLO. 34 00:02:46,000 --> 00:02:50,000 It has introduced the concept of anchor free detection head. 35 00:02:51,000 --> 00:02:52,000 Your lobby. 36 00:02:52,000 --> 00:02:58,000 It is more efficient than the previous version because of its because it uses a larger feature map and 37 00:02:58,000 --> 00:03:01,000 a more efficient convolution neural network. 38 00:03:02,000 --> 00:03:05,000 Okay, so let's move towards the next part. 39 00:03:07,000 --> 00:03:14,000 So, your know, it also uses feature Pyramid networks, which helps to better recognize objects of 40 00:03:14,000 --> 00:03:18,000 different sizes which improve its overall accuracy. 41 00:03:18,000 --> 00:03:25,000 YOLO V8 introduces a user friendly API, allowing users to quickly and easily implement the model in 42 00:03:25,000 --> 00:03:27,000 their applications. 43 00:03:30,000 --> 00:03:37,000 Well, before recording this lecture, I have done a comparison of YOLO V8 with Yolov7 on a custom dataset, 44 00:03:37,000 --> 00:03:40,000 which is license plate detection dataset. 45 00:03:40,000 --> 00:03:42,000 So the dataset is available in roboflow. 46 00:03:42,000 --> 00:03:50,000 I will share the link with as well and the comparison files of YOLO, V7 and YOLO V8 on considering 47 00:03:50,000 --> 00:03:54,000 license plate detection dataset is also attached below this video tutorial. 48 00:03:54,000 --> 00:03:56,000 You can check those files as well. 49 00:03:56,000 --> 00:04:03,000 So the dataset I took is license plate detection dataset and I have done performance comparison of YOLO 50 00:04:03,000 --> 00:04:04,000 V8 with YOLO. 51 00:04:04,000 --> 00:04:08,000 V7 considering this license plate detection dataset. 52 00:04:08,000 --> 00:04:15,000 So the dataset consists of 600 images for training and 64 images for the validation purpose and the 53 00:04:15,000 --> 00:04:23,000 performance metrics, or the how we evaluate with both the models of YOLO, V8 and YOLO. 54 00:04:23,000 --> 00:04:30,000 V7 is based on the training time, how much training time each model take to train for a considering 55 00:04:30,000 --> 00:04:38,000 100 epochs and the performance analysis is done considering the mean average precision with IOU 50 and 56 00:04:38,000 --> 00:04:40,000 IOU 50 to 95. 57 00:04:40,000 --> 00:04:46,000 So the dataset had 601 images for training and 64 images for validation. 58 00:04:46,000 --> 00:04:54,000 The number of epochs was set to 100 to see the performance of YOLO, V8 and Yolov7 model in warmup iteration 59 00:04:54,000 --> 00:04:56,000 means at the starting iterations. 60 00:04:57,000 --> 00:05:00,000 So let's see what the results do we get. 61 00:05:01,000 --> 00:05:08,000 So here are the results I got after training and testing Yolov7 and YOLO V8 models on license plate 62 00:05:08,000 --> 00:05:10,000 detection dataset. 63 00:05:10,000 --> 00:05:13,000 So first I will present a training time comparison. 64 00:05:13,000 --> 00:05:23,000 The training time taken taken by YOLO V8 model to train on 100 epochs is 48.12 minutes, while the training 65 00:05:23,000 --> 00:05:29,000 time taken by YOLO V7 model to train on 100 epochs is 63.06 minutes. 66 00:05:29,000 --> 00:05:37,000 So YOLO V8 trains in lesser time then use yolov7 on the same number of epochs. 67 00:05:37,000 --> 00:05:44,000 So it means usually it takes less training time than YOLO V7, which is a very significant improvement 68 00:05:44,000 --> 00:05:45,000 in YOLO V8. 69 00:05:45,000 --> 00:05:54,000 While the mean average precision which is calculated at a threshold of IOU threshold of 0.5 is with 70 00:05:54,000 --> 00:06:02,000 YOLO v eight gives us a 0.93, which means 93% of mean average precision value, while Yolov7 gives 71 00:06:02,000 --> 00:06:07,000 us the value of 0.817, which is 81.7%. 72 00:06:07,000 --> 00:06:12,000 So in this case as well, you know, we it also performs better than YOLO. 73 00:06:12,000 --> 00:06:19,000 V7 So till now we have seen how the training time as well as for the mean average precision YOLO V8 74 00:06:19,000 --> 00:06:21,000 outperforms YOLO V7. 75 00:06:23,000 --> 00:06:30,000 The mean average precision value obtained when the IOU varies from 0.5 to 0.95. 76 00:06:30,000 --> 00:06:37,000 In the case of zero eight, we got a value of 0.57, while in the case of YOLO V7, we got a value of 77 00:06:37,000 --> 00:06:39,000 0.429. 78 00:06:39,000 --> 00:06:40,000 So in the case of. 79 00:06:40,000 --> 00:06:41,000 In this case as well. 80 00:06:41,000 --> 00:06:44,000 YOLO V8 outperforms Yolov7. 81 00:06:44,000 --> 00:06:48,000 So in all three cases which we have seen now. 82 00:06:48,000 --> 00:06:56,000 YOLO V8 performs better than YOLO V7 so we can easily say that YOLO eight outperforms YOLO V7 and the 83 00:06:56,000 --> 00:06:58,000 previous versions as well. 84 00:07:01,000 --> 00:07:03,000 So here are the snapshots of results. 85 00:07:03,000 --> 00:07:07,000 The CoLab file is also given below this video tutorial. 86 00:07:07,000 --> 00:07:08,000 You can check it as well. 87 00:07:08,000 --> 00:07:17,000 So from the results we can see that if I can show you over here that if you can see that, if this is 88 00:07:17,000 --> 00:07:26,000 the mean average precision 5350 which we got with YOLO v eight is 0.93 and mean average precision with 89 00:07:26,000 --> 00:07:37,000 varies from 50 to 95 is 0.57, while the training time taken is 0.802 hours, which is 48 minutes. 90 00:07:37,000 --> 00:07:42,000 So this is the comparison of YOLO v eight or these are the results of YOLO v eight. 91 00:07:42,000 --> 00:07:48,000 Now let us see the results of YOLO V7 The CoLab file is also attached below this video tutorial. 92 00:07:48,000 --> 00:07:49,000 You can check it as well. 93 00:07:52,000 --> 00:07:52,000 Hi guys. 94 00:07:52,000 --> 00:07:56,000 This is the snapshot of the results which we got with YOLO. 95 00:07:56,000 --> 00:08:00,000 V7 considering license plate detection problem. 96 00:08:00,000 --> 00:08:07,000 So in the previous slide we see the results of V8 model considering license plate detection problem. 97 00:08:07,000 --> 00:08:15,000 In this slide I am discussing the result of Yolov7 model considering license plate detection problem. 98 00:08:15,000 --> 00:08:24,000 So you can see over here the mean average precision value considering IOU of 0.5 we got is 0.817 with 99 00:08:24,000 --> 00:08:25,000 YOLO V7 model. 100 00:08:25,000 --> 00:08:32,000 While in the last slide we have seen that with YOLO V8, the mean average precision value which we got 101 00:08:32,000 --> 00:08:33,000 is 0.93. 102 00:08:34,000 --> 00:08:40,000 So YOLO being the case of YOLO, V7, we have a less value of mean average precision, while in the 103 00:08:40,000 --> 00:08:46,000 case of YOLO V8, we got a better mean average precision than YOLO. 104 00:08:46,000 --> 00:08:50,000 V7 So now let us see the value of mean average precision. 105 00:08:50,000 --> 00:08:57,000 Considering IOU from 0.5 to 0.95 is 0.429. 106 00:08:57,000 --> 00:08:59,000 You can see over here. 107 00:08:59,000 --> 00:09:05,000 So in the case of YOLO V8, we got a mean average precision value of 0.57. 108 00:09:05,000 --> 00:09:10,000 While in the case of YOLO V7 here our value is 0.429. 109 00:09:10,000 --> 00:09:18,000 So the mean average precision considering IOU from 0.5 to 0.95 in the case of YOLO, V7 is less than 110 00:09:18,000 --> 00:09:19,000 YOLO V8. 111 00:09:19,000 --> 00:09:22,000 So in that case, in this case YOLO. 112 00:09:23,000 --> 00:09:31,000 In the case of YOLO V8, we have a higher mean average precision considering IOU of 0.5 and 0.95, while 113 00:09:31,000 --> 00:09:36,000 in the case of Europe V7, our mean average precision value is less than zero. 114 00:09:36,000 --> 00:09:43,000 V8 while that in the case of YOLO V8, the training time taken by the model to train on unrolled epochs 115 00:09:43,000 --> 00:09:51,000 is 48 minutes, while in this case, in the case of YOLO V7 the training time taken by the model to 116 00:09:51,000 --> 00:09:57,000 train on 100 epochs is 1.051 hours, which is 63 minutes. 117 00:09:57,000 --> 00:10:07,000 So YOLO V7 takes a 63 minutes to train on 100 epochs, while YOLO V8 model takes 48 minutes to train 118 00:10:07,000 --> 00:10:09,000 on 100 epochs. 119 00:10:09,000 --> 00:10:13,000 So training time is less in Yolo V8 than in YOLO. 120 00:10:13,000 --> 00:10:21,000 V7 So in your case of YOLO V7, the model takes more training time than YOLO V8 so it means YOLO V8 121 00:10:21,000 --> 00:10:26,000 outperforms YOLO V7 model in all aspects in the case in Europe. 122 00:10:26,000 --> 00:10:33,000 It takes less training time than YOLO, V7 Yolo V3 It gives better mean average precision than YOLO 123 00:10:33,000 --> 00:10:35,000 V7 as well. 124 00:10:38,000 --> 00:10:41,000 The result analysis which we come up with is. 125 00:10:41,000 --> 00:10:45,000 So it basically performs better than Yolov7 model. 126 00:10:45,000 --> 00:10:51,000 YOLO V8, as we have discussed in previous slide, gives us better mean average precision values while 127 00:10:51,000 --> 00:10:52,000 YOLO. 128 00:10:52,000 --> 00:10:55,000 It takes less training time than YOLO V7. 129 00:10:55,000 --> 00:11:02,000 So YOLO V8 gives us maximum value at the expense of reduced time for training and the improvement. 130 00:11:02,000 --> 00:11:06,000 Wearing YOLO is the sensibility of YOLO. 131 00:11:06,000 --> 00:11:07,000 V8 is an important characteristic. 132 00:11:07,000 --> 00:11:15,000 It is created as a framework that works with all prior YOLO iteration makes it easier to switch between 133 00:11:15,000 --> 00:11:21,000 them as so in YOLO V8, we can switch between other versions of YOLO as well. 134 00:11:21,000 --> 00:11:23,000 We can switch with with YOLO V5. 135 00:11:23,000 --> 00:11:26,000 We can also switch between YOLO V3 as well. 136 00:11:26,000 --> 00:11:32,000 So YOLO V8 is the only version which allows us to switch between other versions of YOLO as well, so 137 00:11:32,000 --> 00:11:36,000 we can do a performance comparison with each version as well. 138 00:11:36,000 --> 00:11:38,000 So YOLO. 139 00:11:38,000 --> 00:11:45,000 Because of this, YOLO V8 is the best option for those who wish to benefit from the most recent YOLO 140 00:11:45,000 --> 00:11:49,000 technology while keeping their YOLO models functional. 141 00:11:51,000 --> 00:11:57,000 And on my observation, which I have drawn after doing the comparison between YOLO and YOLO. 142 00:11:57,000 --> 00:12:06,000 V7 So what does have come with the result is that YOLO it takes less training time than YOLO, V7 and 143 00:12:06,000 --> 00:12:14,000 YOLO V8 gives us better mean average precision value then YOLO V7 So in case of YOLO V8, that trade 144 00:12:14,000 --> 00:12:21,000 off between training time as precision is achieved more in YOLO V8 than all the other models of YOLO 145 00:12:22,000 --> 00:12:29,000 and in case of YOLO V8, a new backbone network, a new anchor free detection head and a new loss function 146 00:12:29,000 --> 00:12:31,000 make things much faster. 147 00:12:31,000 --> 00:12:37,000 It blesses the training time and gives us the faster processing and give us the more better results 148 00:12:37,000 --> 00:12:40,000 than all the previous versions of YOLO. 149 00:12:42,000 --> 00:12:44,000 Thank you for watching this video tutorial. 150 00:12:44,000 --> 00:12:47,000 See you all in the next video tutorial. 151 00:12:47,000 --> 00:12:48,000 Till then, bye bye.