1 00:00:03,110 --> 00:00:08,600 The other way, then, is the latest state of the art object detection model that has been developed 2 00:00:08,600 --> 00:00:12,410 by the researchers at Tsinghua University in China. 3 00:00:13,280 --> 00:00:13,820 So YOLO. 4 00:00:13,820 --> 00:00:18,320 It then introduces a novel approach for real time object detection. 5 00:00:19,010 --> 00:00:26,630 So it basically addresses the deficiencies in both post-processing and model architecture that are found 6 00:00:26,630 --> 00:00:32,330 in the earlier YOLO versions, which include Yolo b nine, B eight, Yolo v7. 7 00:00:32,900 --> 00:00:36,710 So what are the problems in previous versions? 8 00:00:36,740 --> 00:00:45,170 So the previous YOLO models, including YOLO, YOLO, V, eight and V7 were using non expression technique 9 00:00:45,170 --> 00:00:46,580 in post-processing. 10 00:00:46,580 --> 00:00:50,450 During inference, I will explain you what is known by expression later on. 11 00:00:50,450 --> 00:00:57,860 But remember the previous models were using non expression techniques in post-processing during inference, 12 00:00:58,340 --> 00:01:03,320 which leads to inefficiencies and increased inference latency time. 13 00:01:04,190 --> 00:01:07,880 So you love it then eliminates the need for the normal expression. 14 00:01:07,880 --> 00:01:11,180 In yolo v ten we are not using the normal expression technique. 15 00:01:11,420 --> 00:01:17,540 Okay, so the non expression technique leads to inefficiencies and increases the inference latency. 16 00:01:18,080 --> 00:01:21,560 So yolo v ten eliminates the need for the normal expression. 17 00:01:21,560 --> 00:01:29,120 And along with this architecture enhancement are made as well in YOLO beta which include optimizing 18 00:01:29,120 --> 00:01:30,620 various model components. 19 00:01:31,540 --> 00:01:38,980 So in short, you can achieve state of the art performance with significantly reduced computational 20 00:01:38,980 --> 00:01:39,550 overhead. 21 00:01:39,700 --> 00:01:47,080 And when wallaby ten is ten state on the benchmark Ms-coco data set, wallaby ten showed shows a superior 22 00:01:47,080 --> 00:01:52,960 accuracy and latency trade off than the other YOLO models, which include YOLO, Be nine, YOLO, V8, 23 00:01:52,960 --> 00:01:54,010 and Yolov7. 24 00:01:54,040 --> 00:01:55,690 So in this tutorial. 25 00:01:56,920 --> 00:02:01,360 We will see what is geology and how geology ten works. 26 00:02:01,360 --> 00:02:06,220 And, uh, we will see what architecture enhancements are made in geology ten. 27 00:02:06,220 --> 00:02:11,680 And we will do a performance comparison of we ten with other Yolov2 models as well. 28 00:02:11,680 --> 00:02:14,170 And that's all what we will cover in this tutorial. 29 00:02:14,170 --> 00:02:15,460 So let's get started. 30 00:02:15,460 --> 00:02:22,060 So YOLO ten is the real time state of the art object detection model introduced in the paper YOLO v 31 00:02:22,060 --> 00:02:24,370 ten Real-Time End to End Object Detection. 32 00:02:24,370 --> 00:02:26,560 So this paper is available online. 33 00:02:26,560 --> 00:02:31,240 I have just added a snapshot of this paper so you can review this complete paper as well. 34 00:02:31,240 --> 00:02:35,050 In this tutorial I will try to present the crux of this paper. 35 00:02:35,050 --> 00:02:36,790 What is inside this paper. 36 00:02:36,790 --> 00:02:39,640 So I will just try to present the crux of this paper. 37 00:02:39,640 --> 00:02:46,180 So YOLO ten is released in May 2024 is a new advancement in the field of real time object detection. 38 00:02:46,180 --> 00:02:52,000 Basically, YOLO ten tries to address the issues that are faced by the, uh, previous models. 39 00:02:52,000 --> 00:02:55,510 Like what enhancements are made in YOLO ten plus. 40 00:02:55,840 --> 00:03:01,450 Our model architecture enhancements are made in YOLO plus YOLO and addresses the post-processing issues 41 00:03:01,570 --> 00:03:02,020 like the. 42 00:03:02,020 --> 00:03:04,720 It eliminates the need for the random expression. 43 00:03:07,350 --> 00:03:07,590 But. 44 00:03:08,510 --> 00:03:10,310 So what is YOLO we then? 45 00:03:10,310 --> 00:03:12,770 YOLO v ten is a cutting edge. 46 00:03:13,220 --> 00:03:18,620 Computer vision architecture designed for real time object detection, and it is built upon the advancements 47 00:03:18,620 --> 00:03:19,970 of its predecessor. 48 00:03:20,600 --> 00:03:23,660 Okay, Yolo v ten model achieves a higher mean average compare. 49 00:03:23,810 --> 00:03:31,070 Mean average precision compared to earlier modular models such as V9, Yolo V8, Yolo v7 when benchmarked 50 00:03:31,070 --> 00:03:33,020 against the Ms-coco data set. 51 00:03:33,020 --> 00:03:38,330 So you can see over here, um, I will explain it in the Dail as well, but you can see over here this 52 00:03:38,330 --> 00:03:41,330 is the accuracy average precision, and this is the latency. 53 00:03:41,330 --> 00:03:46,010 So we can see an accuracy and uh latency and accuracy trade off over here. 54 00:03:46,010 --> 00:03:47,930 And you can see this red line. 55 00:03:47,960 --> 00:03:55,910 So this is uh you can see over here YOLO V ten outperforms all the previous models in terms of accuracy 56 00:03:55,910 --> 00:03:58,580 as well as in terms of latency as well. 57 00:03:58,580 --> 00:04:00,530 So what is latency basically? 58 00:04:00,620 --> 00:04:05,600 Uh, latency is basically the time taken to do object detection on an input image. 59 00:04:05,600 --> 00:04:06,830 So this is latency. 60 00:04:06,830 --> 00:04:12,170 Latency is basically the time which is taken to do object detection on an input image. 61 00:04:12,170 --> 00:04:19,580 So we can say that YOLO with ten uh takes less time as compared to other YOLO models like you can see 62 00:04:19,580 --> 00:04:20,330 over here. 63 00:04:20,600 --> 00:04:26,870 Uh, from here we can see that YOLO ten definitely takes very much less time as compared to other YOLO 64 00:04:26,870 --> 00:04:33,230 models like, uh, it, uh, quickly does object detection on an input image or an input frame than 65 00:04:33,230 --> 00:04:34,250 other YOLO models. 66 00:04:34,250 --> 00:04:40,640 Plus it you can see over here, uh, YOLO ten gives good accuracy like you can see over here in comparison 67 00:04:40,640 --> 00:04:42,350 to other YOLO retirement. 68 00:04:42,440 --> 00:04:44,000 Uh, YOLO models. 69 00:04:44,450 --> 00:04:47,540 So, uh, that is all, uh. 70 00:04:48,690 --> 00:04:53,880 So YOLO with an object or various strategies to tackle the limitations of previous ruler models. 71 00:04:53,880 --> 00:04:58,110 Earlier neural models rely on non expression for post-processing during inference. 72 00:04:58,110 --> 00:05:05,430 I told you that uh, earlier models rely on the non expression for technique uh, during post-processing, 73 00:05:05,430 --> 00:05:10,860 uh, while doing inference which leads to inefficiencies and increase the inference latency. 74 00:05:10,860 --> 00:05:17,970 So the basically if I use Non-max suppression technique it post-processing uh during inference, uh, 75 00:05:17,970 --> 00:05:21,060 basically this will lead to uh, more latency time. 76 00:05:21,060 --> 00:05:25,200 Like it will take more time to process, uh, to do object detection on input image. 77 00:05:25,500 --> 00:05:33,180 Uh, and plus this will also compromise on the accuracy as well, which basically comes with inefficiencies. 78 00:05:33,240 --> 00:05:38,340 Uh, due to uh non expression comes with some inefficiencies to address these limitations. 79 00:05:38,340 --> 00:05:42,300 YOLO we then comes with a consistent view assignment strategy. 80 00:05:42,300 --> 00:05:48,060 So Yolo v ten basically adopts a consistent dual assignment strategy which eliminates the need for non 81 00:05:48,120 --> 00:05:50,310 expression during inference. 82 00:05:50,310 --> 00:05:53,640 And it significantly reduces the inference latency. 83 00:05:53,640 --> 00:05:59,190 So now you can see in yolo v ten we have eliminated or removed uh non expression technique. 84 00:05:59,190 --> 00:06:04,740 So by removing normal expression technique you can see that our latency time has reduced or reduced 85 00:06:04,740 --> 00:06:05,310 a bit. 86 00:06:05,550 --> 00:06:08,250 Uh our latency time has reduced significantly. 87 00:06:08,250 --> 00:06:13,530 Plus you can see that our accuracy has improved like YOLO outperforms all the other object detection 88 00:06:13,530 --> 00:06:14,220 models. 89 00:06:15,400 --> 00:06:17,020 So long. 90 00:06:17,080 --> 00:06:24,850 Expression technique is evaluated during inference in your log10, and it significantly reduces latency, 91 00:06:24,850 --> 00:06:28,300 the increased latency, while retaining competitive performance. 92 00:06:28,780 --> 00:06:36,010 So Jilawatan incorporates efficiency accuracy driven strategy, which involves optimizing various components 93 00:06:36,010 --> 00:06:40,540 of the model to minimize computational overhead enhance performance. 94 00:06:40,540 --> 00:06:45,340 So Yolov3 ten basically operates efficiency accuracy driven design strategy. 95 00:06:45,340 --> 00:06:49,450 So when we adopt efficiency accuracy driven design strategy. 96 00:06:49,450 --> 00:06:53,410 So we are optimizing various components in the YOLO model. 97 00:06:53,410 --> 00:06:58,330 We will discuss what components we are optimizing, what architecture enhancements are made in YOLO 98 00:06:58,510 --> 00:07:00,190 and later as we go ahead. 99 00:07:00,190 --> 00:07:06,310 So now you can see that as we adopt efficiency accuracy driven design strategy, uh, and we enhance 100 00:07:06,310 --> 00:07:10,930 various component of the model so that we can reduce computational overhead and enhance performance. 101 00:07:10,930 --> 00:07:16,360 And we can see in the graph as well, like you can see over here, you know, we didn't use uh, models 102 00:07:16,360 --> 00:07:21,550 has less number of parameters as compared to other models like these are in billions. 103 00:07:21,550 --> 00:07:22,930 Parameters are in millions. 104 00:07:22,930 --> 00:07:26,620 So now you can see over here this is the red color is for YOLO ten. 105 00:07:26,620 --> 00:07:31,330 And you can see that YOLO ten has uses less number of parameters. 106 00:07:31,570 --> 00:07:31,960 Uh. 107 00:07:32,940 --> 00:07:39,510 As compared to other YOLO models, and YOLO Eden has better accuracy as compared to all the other YOLO 108 00:07:40,770 --> 00:07:41,460 models. 109 00:07:43,010 --> 00:07:48,710 So now you can see that, uh, we have minimized by adopting efficiency, accuracy driven design strategy. 110 00:07:48,710 --> 00:07:50,780 We have minimized the computation overhead. 111 00:07:50,780 --> 00:07:57,320 And unlike less number of parameters are used and we have enhanced the model performance as well. 112 00:08:00,890 --> 00:08:06,830 So like we are discussing about Non-expression, we have evaluated non-max suppression during, in, 113 00:08:06,830 --> 00:08:09,380 uh, in post-processing, during inference. 114 00:08:09,380 --> 00:08:14,930 So what we are saying that, uh, YOLO eliminates the need for non-expert expression during inference. 115 00:08:14,930 --> 00:08:16,730 So what is Non-gm expression? 116 00:08:18,470 --> 00:08:24,380 So non expression is basically a post-processing technique used in object detection to remove the redundant 117 00:08:24,380 --> 00:08:26,210 or overlapping bounding boxes. 118 00:08:26,480 --> 00:08:29,840 Uh, the main aim of normal expression is to retain only the bounding boxes. 119 00:08:29,840 --> 00:08:33,740 Higher confidence score and the bounding boxes with lower confidence score are suppressed or removed. 120 00:08:34,070 --> 00:08:39,710 So now you can see that we have detected, uh, truck or a car in this image, like you can see over 121 00:08:39,710 --> 00:08:40,100 here. 122 00:08:40,100 --> 00:08:46,340 And you can see that after doing object detection with YOLO, we then, uh, what output we get is that, 123 00:08:46,340 --> 00:08:51,770 uh, we have that object detected, like, you can see over here, um, the truck, but you can see 124 00:08:51,770 --> 00:08:54,410 that we have multiple bounding boxes as well. 125 00:08:54,410 --> 00:09:01,190 So we will be we use non expression technique in earlier YOLO models like YOLO, YOLO, yolo v7 used 126 00:09:01,190 --> 00:09:05,630 non extrapolated non max suppression technique during inference. 127 00:09:05,810 --> 00:09:09,620 Uh so that they can remove the redundant or overlapping bounding boxes. 128 00:09:09,620 --> 00:09:11,270 So now you can see there is only one truck. 129 00:09:11,270 --> 00:09:15,050 So there should be only one bounding box like you can see over here. 130 00:09:15,530 --> 00:09:15,770 So. 131 00:09:16,970 --> 00:09:22,940 A little YOLO models by using non-max suppression techniques so that they can remove redundant or overlapping 132 00:09:22,940 --> 00:09:26,990 bounding boxes for the bounding box, which have the highest confidence score among all these bounding 133 00:09:26,990 --> 00:09:28,070 boxes will be retained. 134 00:09:28,070 --> 00:09:32,330 Like this morning box will have the highest confidence score than all the other bounding boxes, so 135 00:09:32,330 --> 00:09:33,080 it is retained. 136 00:09:33,290 --> 00:09:34,070 So. 137 00:09:34,820 --> 00:09:40,310 So previous models were using norm expression technique in during inference so that they can emulate 138 00:09:40,310 --> 00:09:42,830 the, uh, overlapping bounding boxes. 139 00:09:42,830 --> 00:09:45,230 But this is uh, leading to inefficiencies. 140 00:09:45,230 --> 00:09:48,830 And it was also increasing latency as well. 141 00:09:48,830 --> 00:09:53,030 So in rollup we then we have eliminated the need for the norm expression. 142 00:09:54,790 --> 00:09:58,120 So here you can see we will discuss how you have written words. 143 00:09:58,120 --> 00:10:02,470 You also written introduces a novel training strategy and architecture enhancement. 144 00:10:02,500 --> 00:10:05,380 Let's discuss the main component of all your written work. 145 00:10:05,440 --> 00:10:11,110 So yellow written produces an ms3 training strategy with new label assignments. 146 00:10:11,110 --> 00:10:13,000 So this is a snapshot from the paper I have. 147 00:10:13,000 --> 00:10:15,760 This app added our tracks over here. 148 00:10:15,760 --> 00:10:17,320 What is inside the paper. 149 00:10:17,620 --> 00:10:24,790 So YOLO written basically docs on novel training strategy and it provides some architecture enhancements 150 00:10:24,790 --> 00:10:27,970 as well or introduces some architecture enhancement as well. 151 00:10:28,240 --> 00:10:33,370 So like I told you that, uh, YOLO imitates the need for non-max expression. 152 00:10:33,370 --> 00:10:41,710 So traditional YOLO models like YOLO, benign yolo V8 employ 1 to 1 one to many assignments strategy 153 00:10:41,710 --> 00:10:44,950 during training which make it necessary to use norm expression. 154 00:10:44,950 --> 00:10:51,760 So the earlier models like YOLO benign V8 were using uh, norm expression technique. 155 00:10:51,760 --> 00:10:58,090 Or you can see that the earlier models were adopting one to many assignment strategy, which make it 156 00:10:58,090 --> 00:11:02,950 necessary to use norm expression during inference, which like we use norm expression. 157 00:11:02,950 --> 00:11:08,740 I told you that to filter out redundant or overlapping bounding boxes, which leads to inefficiencies 158 00:11:08,740 --> 00:11:10,990 and increase inference latency. 159 00:11:11,350 --> 00:11:14,650 Okay, so earlier YOLO models adopt one to many assignment strategy. 160 00:11:14,980 --> 00:11:21,730 So YOLO v ten adopts a new label assignment strategy that incorporates one to many like YOLO, and uses 161 00:11:21,730 --> 00:11:25,870 the one to many assignment strategy and 1 to 1 matching approaches. 162 00:11:25,870 --> 00:11:31,450 Like now you will be thinking that YOLO written observes one to many and 1 to 1 matching approaches 163 00:11:31,450 --> 00:11:31,840 like. 164 00:11:31,840 --> 00:11:38,410 But by thinking that in one to many assignment strategy they are using in non expression during inference 165 00:11:38,410 --> 00:11:40,630 to filter out predicted bounding boxes. 166 00:11:40,630 --> 00:11:46,990 But as we adopt one to many and 1 to 1 matching approaches, we will not be using non expression during 167 00:11:46,990 --> 00:11:47,560 inference. 168 00:11:47,560 --> 00:11:49,510 So what is 1 to 1 matching. 169 00:11:49,840 --> 00:11:51,400 So in 1 to 1 matching the. 170 00:11:51,400 --> 00:11:57,250 Our model assigns a single prediction to each ground truth instance eliminates the need for non expression. 171 00:11:57,250 --> 00:12:01,960 So in 1 to 1 uh matching we don't use non expression okay. 172 00:12:02,470 --> 00:12:08,920 But uh this results in weaker supervision or could be causing suboptimal accuracy and slower convergence. 173 00:12:08,920 --> 00:12:12,790 So but uh if we 1 to 1 matching we are not using norm expression. 174 00:12:12,790 --> 00:12:16,450 But uh this results in some compromise on accuracy. 175 00:12:16,630 --> 00:12:23,920 But in 1 to 1, one to many assignments, uh, although it provides richer supervisory signals or signals 176 00:12:23,920 --> 00:12:26,950 but uh required non expression per inference. 177 00:12:26,950 --> 00:12:31,180 So a one to many assignments we will be using non expression for inference. 178 00:12:31,180 --> 00:12:33,340 But this leads to uh. 179 00:12:34,610 --> 00:12:36,410 At latency like increased latency. 180 00:12:36,410 --> 00:12:41,570 So in 1 to 1 matching we don't use non-expression, but we are compromising accuracy. 181 00:12:41,570 --> 00:12:45,470 But in one to many, many segments we are not compromising on accuracy. 182 00:12:45,470 --> 00:12:50,330 We are using non-expression in inference, but this leads to increased latency. 183 00:12:50,600 --> 00:12:58,040 So how YOLO we can address this issue like not using non-expression plus not compromising on accuracy 184 00:12:58,040 --> 00:13:01,310 and also reducing the inference latency. 185 00:13:03,640 --> 00:13:08,830 Okay, so you love it and cleverly combine these strategies by introducing an additional 1 to 1 head. 186 00:13:08,830 --> 00:13:12,850 So basically, you know, we tend to use introducing an additional one one head. 187 00:13:13,060 --> 00:13:13,900 And. 188 00:13:15,350 --> 00:13:20,060 So mirroring the original one to many branch structure and optimization objectives. 189 00:13:20,330 --> 00:13:25,430 During start training, both heads are jointly optimized labeling that requires supervision from one 190 00:13:25,430 --> 00:13:26,780 to many assignments. 191 00:13:26,810 --> 00:13:30,110 So now this is embodied during the YOLO model. 192 00:13:30,110 --> 00:13:35,150 Initializes only 1 to 1 head, thus bypassing the need for norm expression. 193 00:13:35,150 --> 00:13:42,230 So while doing inference, YOLO with ten only adopts 1 to 1 approach, and it bypasses the need for 194 00:13:42,230 --> 00:13:47,480 norm expression and achieve high efficiency without adding additional inference cost. 195 00:13:47,480 --> 00:13:49,460 So this is how user written works. 196 00:13:51,090 --> 00:13:55,440 So another thing like consistent matching that is like we can see all the details in the paper. 197 00:13:55,440 --> 00:13:59,670 But the crux of this is that a key component of the dual assignment strategy. 198 00:13:59,670 --> 00:14:07,080 So as I told you, adopts a dual assignment strategy that uh, combines uh 1 to 1 matching and one to 199 00:14:07,080 --> 00:14:09,990 many assignments like you can see over here. 200 00:14:10,290 --> 00:14:15,180 So now you can see that dual label assignment strategy of one to many and 1 to 1 approach as well. 201 00:14:15,180 --> 00:14:21,390 But now you can see that after one to many and 1 to 1 they will be spent matching metric. 202 00:14:21,510 --> 00:14:27,060 So a key component of the dual label assignment strategy that we are adopting in YOLO and and so YOLO 203 00:14:27,060 --> 00:14:31,980 we then uh, adopted a novel approach which is called dual label assignment strategy. 204 00:14:31,980 --> 00:14:35,040 And here we have the complete architecture like how it works. 205 00:14:35,040 --> 00:14:39,840 So you can see a key component in new label assignment strategy is persistent matching metric. 206 00:14:39,990 --> 00:14:45,480 A few key component of that new label assignment strategy is the persistent matching metrics, which 207 00:14:45,480 --> 00:14:52,770 is used to evaluate the concordance between the prediction and ground truth instances like calculate 208 00:14:52,770 --> 00:14:58,200 the matching between the prediction and the ground truth strategy so you can see how close they are 209 00:14:58,200 --> 00:15:05,340 or how far away they are, like how close our predictions are with from, uh, ground truth and or how 210 00:15:05,340 --> 00:15:07,080 far our predictions are from the ground. 211 00:15:09,080 --> 00:15:12,590 So now you can see what architecture enhancements are made in your weekend. 212 00:15:12,620 --> 00:15:19,700 So the component of YOLO models are traditionally the components of YOLO model consist of the stem downsampling 213 00:15:19,700 --> 00:15:22,430 layer stages with basic building blocks, and the head. 214 00:15:22,970 --> 00:15:29,390 Uh Yolo focusing on optimizing the other three parts to enhance, uh, efficiency. 215 00:15:29,390 --> 00:15:34,100 So uh, we will be ten introduces basically lightweight classification had in it. 216 00:15:34,100 --> 00:15:38,810 And our light class lightweight classification head is designed to reduce the computational dependency 217 00:15:38,810 --> 00:15:41,810 in ensuring that a model operates more efficiently. 218 00:15:42,440 --> 00:15:44,600 Sparkle channel decoupled downsampling. 219 00:15:44,630 --> 00:15:51,200 So, uh, in YOLO, we dance party channel decoupled downsampling is applied to the Wise feature extraction, 220 00:15:51,200 --> 00:15:55,790 making uh the process more efficient and rank guided block design. 221 00:15:55,790 --> 00:16:01,670 The rank guided block design further streamlines the architecture, enhancing overall efficiency and 222 00:16:01,670 --> 00:16:02,960 large kernel convolution. 223 00:16:02,960 --> 00:16:07,670 The large kernel convolution is utilized to improve the model's capacity to capture detailed features, 224 00:16:07,670 --> 00:16:13,430 and the effective partial self-attention module most aggressive with minimal computational cost. 225 00:16:13,460 --> 00:16:16,250 So these are the architecture enhancement that are made in YOLO. 226 00:16:18,030 --> 00:16:23,520 So now we are doing a YOLO v ten performance comparison with other baseline models like YOLO, YOLO 227 00:16:23,520 --> 00:16:26,460 v ten comparison to other baseline models like v eight. 228 00:16:26,520 --> 00:16:28,260 So you can skip this. 229 00:16:28,290 --> 00:16:33,810 We are not, uh, using YOLO we ten D model in comparison because currently we are doing comparison 230 00:16:33,810 --> 00:16:36,060 with yolo v eight models. 231 00:16:36,060 --> 00:16:42,690 And uh, we are skipping this and we will be using yolo v ten and yolo with n s, YOLO with ten M, 232 00:16:42,690 --> 00:16:47,190 YOLO with n, l and x v x model for the comparison. 233 00:16:48,210 --> 00:16:54,120 In comparison to other baseline models like, there will be ten more straight improvements of 1.2%, 234 00:16:54,120 --> 00:17:01,290 1.4%, 0.5, 0.3 0.5% in average precision like you can see over here. 235 00:17:01,380 --> 00:17:03,390 Here we have the validation average precision. 236 00:17:03,390 --> 00:17:06,990 You can compare the yellow with ten, uh model average precision. 237 00:17:06,990 --> 00:17:08,610 And this yellow eight. 238 00:17:08,820 --> 00:17:14,670 And here you can see your S model average precision with yolo V8 S model over here. 239 00:17:14,670 --> 00:17:20,100 And so after doing this comparison you will see that your W ten demonstrate improvements by the numbers 240 00:17:20,100 --> 00:17:28,080 provided over here in average precision with 28%, 36%, 41%, 44%, 57 fewer parameters. 241 00:17:28,080 --> 00:17:34,560 Like you can see the parameters found in million by end users 2.3 in uh, yellow ten and model uses 242 00:17:34,560 --> 00:17:35,370 nano models. 243 00:17:35,370 --> 00:17:41,640 You can use this 2.3 million parameters, and Yolo V8 nano models uses 3.2 million parameters for definitely 244 00:17:41,790 --> 00:17:47,310 yellow B ten model uses less number of parameters, and YOLO with ten small model is 7.2 million parameters 245 00:17:47,310 --> 00:17:53,370 that YOLO small model uses 11.2 million parameters, so definitely using one model with this less number 246 00:17:53,370 --> 00:17:54,240 of parameters. 247 00:17:55,660 --> 00:18:05,890 And you can see that, uh, uh, with 28%, 36%, 41%, 44%, and 57% fewer minorities and eight model 248 00:18:06,280 --> 00:18:20,140 land uses, and 23%, 40, 24%, 25%, 27% 30% fewer calculation and 70% 65%, 50%, 41%, 37% lower 249 00:18:20,140 --> 00:18:20,710 latencies. 250 00:18:20,710 --> 00:18:25,300 Like you can see over here, latency in milliseconds with ten is and models. 251 00:18:25,300 --> 00:18:33,280 Nano models have 1.84, and the other V8 and nano model has 6.16 with ten, small model has 2.49 and 252 00:18:33,280 --> 00:18:33,820 Yolov5. 253 00:18:33,850 --> 00:18:36,640 A small model has 7.07. 254 00:18:36,640 --> 00:18:42,250 So you can see that Yolo v ten has a low latency as compared to the other. 255 00:18:42,250 --> 00:18:42,640 We're. 256 00:18:44,980 --> 00:18:49,630 So in conclusion, your dividend represents a significant investment in real time object detection and 257 00:18:49,990 --> 00:18:52,810 state of the art performance in terms of speed and accuracy. 258 00:18:52,810 --> 00:18:58,600 So Yolov3 ten introduces NMS speed training and it adopts an efficiency accuracy driven model design 259 00:18:58,600 --> 00:18:59,080 strategy. 260 00:18:59,080 --> 00:19:04,090 In efficiency accuracy driven model design strategy, we optimized various model components so that 261 00:19:04,090 --> 00:19:08,950 we can reduce the number of parameters and get more accurate results. 262 00:19:08,950 --> 00:19:14,530 So you do return includes improves accuracy while reducing computational redundancy and latency. 263 00:19:14,530 --> 00:19:20,110 Competitive analysis against baseline models like if we compare YOLO with ten model with eight Yolov7, 264 00:19:20,560 --> 00:19:26,680 our European R demonstrates superior performance in average precision, parameter efficiency and inference 265 00:19:26,680 --> 00:19:27,010 speed. 266 00:19:27,280 --> 00:19:29,020 Thank you for watching this tutorial.