1
00:00:02,740 --> 00:00:09,220
Wallaby nine is the new computer vision object detection model released by Chenyang Wang and his team

2
00:00:09,220 --> 00:00:11,860
on 21st February 2024.

3
00:00:11,890 --> 00:00:17,530
In this video tutorial, I will give you a detailed overview of the architectural improvements that

4
00:00:17,530 --> 00:00:23,500
are introduced in YOLO V9 model on 21st February 2024.

5
00:00:23,530 --> 00:00:31,750
Chenyang Wang and his team released a paper titled Yolo v9 Learning What You Want to Learn using programmable

6
00:00:31,750 --> 00:00:33,340
Gradient Information.

7
00:00:33,340 --> 00:00:39,880
In this paper, a new computer vision model architecture is introduced, which is YOLO v9.

8
00:00:39,970 --> 00:00:46,810
The source code was made available on GitHub, allowing anyone to train their own YOLO v9 model so you

9
00:00:46,810 --> 00:00:53,050
can fine tune or train the YOLO v9 model on any of their custom data set, and you can fine tune the

10
00:00:53,050 --> 00:00:55,240
YOLO v9 model as per your requirements.

11
00:00:55,240 --> 00:01:01,150
You can fine tune the YOLO v9 model on license plate data set, and you can do license plate detection

12
00:01:01,150 --> 00:01:02,140
and recognition.

13
00:01:02,140 --> 00:01:07,510
You can fine tune the YOLO v9 model on potholes data set, and you can do potholes detection.

14
00:01:07,510 --> 00:01:12,910
You can also do pen book detection by fine tuning the YOLO v9 model on pen book data set.

15
00:01:12,910 --> 00:01:16,480
And in this way, you can perform multiple, uh, tasks.

16
00:01:16,480 --> 00:01:22,210
You can also do personal protective equipment detection by training or fine tuning the YOLO v9 model

17
00:01:22,210 --> 00:01:23,740
on PPE data set.

18
00:01:24,840 --> 00:01:27,900
So from the results provided in the paper.

19
00:01:28,810 --> 00:01:34,570
The wallaby nine achieves higher mean average precision than existing models, which include Yolo,

20
00:01:34,570 --> 00:01:42,580
V8, Yolov7, YOLO, Yolo v5 when benchmarked against the Mscoco validation data set.

21
00:01:43,270 --> 00:01:47,800
So Yolo v nine paper introduces two new architectures YOLO v nine and Glenn.

22
00:01:47,800 --> 00:01:55,000
So both of these uh YOLO v nine weights and Glenn architecture weights are available on the GitHub repository

23
00:01:55,000 --> 00:01:56,050
of Yolo v nine.

24
00:01:56,050 --> 00:01:59,080
You can try YOLO v nine weights and Glenn weights.

25
00:01:59,080 --> 00:02:05,170
And you can do object detection on images, videos or on the live webcam feed as well.

26
00:02:06,020 --> 00:02:11,600
So YOLO v nine introduces two new architectures, Yolo v nine and Glenn, and both of these models,

27
00:02:11,600 --> 00:02:17,330
weights YOLO nine and Glenn weights are available on the YOLO v nine GitHub repository, and you can

28
00:02:17,330 --> 00:02:23,900
use those weights and do object detection on images, videos, and on the live webcam feed going ahead.

29
00:02:23,900 --> 00:02:31,370
Jinyao Wang and his team have also developed YOLO v4, Yolo R, and Yolov7, so YOLO R, Yolov7 and

30
00:02:31,370 --> 00:02:37,370
YOLO v4 have also been developed by the Chen Yang Wang and his team, while Yolov5 and Yolo v eight

31
00:02:37,610 --> 00:02:39,770
are developed by the Ultralytics.

32
00:02:40,810 --> 00:02:48,250
You know, V9 is an advancement from Yolov7, as the previous version of YOLO model, Yolov7 was also

33
00:02:48,250 --> 00:02:50,410
developed by the Shenyang Wang and his team.

34
00:02:50,410 --> 00:02:54,190
So YOLO v9 is basically an advancement from Yolov7.

35
00:02:54,190 --> 00:03:01,360
What, uh, what uh, basically improvements that is made from Yolov7 in YOLO v9 let's discuss those

36
00:03:01,360 --> 00:03:02,620
in Yolov7.

37
00:03:02,620 --> 00:03:06,760
A significant progress is made in terms of optimizing the training process.

38
00:03:06,760 --> 00:03:09,520
That is why it is called trainable bag of freebies.

39
00:03:09,520 --> 00:03:16,630
So if you read the title of the Yolov7 paper, so it is Yolov7 Trainable Bag of freebies and why we

40
00:03:16,630 --> 00:03:24,010
called Yolov7 trainable bag of Freebies, because in yolov7, a significant improvement is made in terms

41
00:03:24,010 --> 00:03:30,940
of optimizing the training process, which, uh, results in harnessing the training efficiency to boost

42
00:03:30,940 --> 00:03:35,530
the object detection model accuracy without adding to the inference cost.

43
00:03:36,900 --> 00:03:42,240
So in, uh, in yolov7 we have optimized the training process.

44
00:03:42,240 --> 00:03:49,320
But the in yolov7 we does not the authors do not specifically address the problem of information loss.

45
00:03:49,320 --> 00:03:55,710
So in Yolov7 the problem of information loss is not addressed, although they have optimized the training

46
00:03:55,710 --> 00:03:56,400
process.

47
00:03:56,400 --> 00:04:02,850
So Yolov7 does not specifically address the problem of information loss during the input data feed forward

48
00:04:02,850 --> 00:04:07,140
process, which is a challenge known as information bottleneck.

49
00:04:07,140 --> 00:04:13,680
So the information loss problem is basically a challenge which they have called as information bottleneck.

50
00:04:13,680 --> 00:04:17,160
So the issue arises from downscaling operations in the network.

51
00:04:17,160 --> 00:04:23,700
So we are the information of the problem of information loss occurs when we, uh perform downscaling

52
00:04:23,700 --> 00:04:25,260
operations in the network.

53
00:04:25,260 --> 00:04:32,580
And as a result of this, an important input data which we have passed to the model can be diluted or

54
00:04:32,580 --> 00:04:33,240
removed.

55
00:04:34,660 --> 00:04:41,290
So to address the issue of information loss, which we have called as information bottleneck, there

56
00:04:41,290 --> 00:04:42,190
exists some solution.

57
00:04:42,190 --> 00:04:49,390
So there are already some solutions which can solve the issue of information loss or information bottleneck,

58
00:04:49,390 --> 00:04:54,040
which include reversible architectures, mask modeling and deep supervision.

59
00:04:54,520 --> 00:05:03,370
So these solutions also help us to uh, uh, reduce the issue of information loss or information bottleneck.

60
00:05:03,370 --> 00:05:11,230
But all of these solutions have some drawbacks by which we have to, uh, which can also result in,

61
00:05:11,230 --> 00:05:14,170
uh, minimizing the accuracy of our object detection model.

62
00:05:14,650 --> 00:05:21,820
So to overcome the issue of information loss information bottleneck, uh, new uh enhancement is being

63
00:05:21,820 --> 00:05:24,790
introduced in the Yolo VI nine paper, which is.

64
00:05:25,970 --> 00:05:31,520
Two different innovative approaches have been used in the following nine papers to address the issue

65
00:05:31,520 --> 00:05:37,880
of information loss, which include programmable gradient information and the Generalized Efficient

66
00:05:37,880 --> 00:05:40,220
Layer Aggregation network.

67
00:05:40,220 --> 00:05:46,700
So two different approaches, which include programmable gradient information and the Generalized Efficient

68
00:05:46,700 --> 00:05:53,090
Layer aggregation network is being used to tackle the information bottleneck or information loss problem

69
00:05:53,090 --> 00:05:53,720
directly.

70
00:05:53,720 --> 00:05:59,750
And if we solve this information bottleneck or information loss problem, this help us to improve the

71
00:05:59,750 --> 00:06:04,370
accuracy and efficiency of object detection model further as well.

72
00:06:05,740 --> 00:06:07,510
Doing the Yolov7 paper.

73
00:06:07,510 --> 00:06:14,350
The authors have optimized the training process, and in the YOLO v9 they have solved the, uh, information

74
00:06:14,350 --> 00:06:19,960
loss issue or information bottleneck issue by adding or introducing two different approaches, which

75
00:06:19,960 --> 00:06:25,030
include programmable gradient information and the generalized efficient layer aggregation network.

76
00:06:25,030 --> 00:06:30,160
So these two approaches will be used to tackle the information bottleneck or information loss problem.

77
00:06:30,160 --> 00:06:36,550
And it will result in further increase of accuracy and efficiency of the object detection model.

78
00:06:37,780 --> 00:06:43,090
So I have just given you a quick interview or a detailed interview about the Yolov5 model.

79
00:06:43,090 --> 00:06:52,450
So what advancement and inaccuracy or architecture are made in Yolov5 nine, and how, uh, how Yolov5

80
00:06:52,450 --> 00:06:54,280
nine outperforms other models?

81
00:06:54,280 --> 00:06:55,750
Let us see that as well.

82
00:06:56,510 --> 00:07:02,810
So Eulabee nine introduces architecture enhancement, which sets your V9 apart from the other object

83
00:07:02,810 --> 00:07:03,830
detection models.

84
00:07:03,830 --> 00:07:04,850
So.

85
00:07:05,910 --> 00:07:10,260
Let's look at what architectural improvements that are made in YOLO v9.

86
00:07:10,260 --> 00:07:16,980
So YOLO v9 incorporates advancements like programmable gradient formation and the Generalized Efficient

87
00:07:16,980 --> 00:07:18,600
Layer aggregation network.

88
00:07:18,600 --> 00:07:21,660
So these two architecture improvements are made in YOLO.

89
00:07:21,660 --> 00:07:27,060
V9 one is programmable gradient information and the other is generalized efficient layer aggregation

90
00:07:27,060 --> 00:07:27,660
network.

91
00:07:27,660 --> 00:07:30,690
So what does programmable gradient information do.

92
00:07:30,720 --> 00:07:36,030
Programmable gradient information prevents data loss or information loss during gradient updates.

93
00:07:36,030 --> 00:07:39,600
And what does generalized efficient layer aggregation network do?

94
00:07:39,630 --> 00:07:45,510
Generalized efficient layer aggregation network optimizes lightweight models through gradient path planning.

95
00:07:45,510 --> 00:07:49,620
So these are the two uh architecture improvements made in YOLO v9.

96
00:07:49,620 --> 00:07:51,390
And what does they perform?

97
00:07:51,390 --> 00:07:52,380
I have already explained.

98
00:07:53,700 --> 00:07:59,040
So the inclusion of a programmable gradient information and the adaptable generalized efficient layer

99
00:07:59,040 --> 00:08:05,790
aggregation network into the architecture of YOLO nine not only boosts the model learning capabilities,

100
00:08:05,790 --> 00:08:09,450
but it also guarantees the preservation of wider information.

101
00:08:09,450 --> 00:08:15,600
So if we incorporate programmable gradient information and generalized efficient layer aggregation network

102
00:08:15,600 --> 00:08:20,730
into the YOLO v nine, this will not only boost the model learning capabilities, but it will also make

103
00:08:20,730 --> 00:08:26,850
sure that no information is lost throughout the detection process, so which results in increase in

104
00:08:26,850 --> 00:08:29,160
accuracy and performance as well.

105
00:08:30,550 --> 00:08:37,180
So we can say that in YOLO, v nine is basically centered around tackling the issue that arises from

106
00:08:37,180 --> 00:08:39,220
information loss in deep neural networks.

107
00:08:39,220 --> 00:08:45,520
So in YOLO nine, the authors Zhenya Wang and his team have addressed the issue of information loss

108
00:08:45,610 --> 00:08:48,490
by using PGI and Glen architecture.

109
00:08:50,220 --> 00:08:56,070
So you using the YOLO model you can do object detection train object detection model on custom data

110
00:08:56,070 --> 00:08:56,700
set.

111
00:08:56,700 --> 00:09:02,700
But we cannot perform segmentation, classification and pose estimation task with YOLO v9 currently.

112
00:09:02,700 --> 00:09:08,970
So as as I'm recording this tutorial on 19th of March, currently you cannot perform segmentation,

113
00:09:08,970 --> 00:09:13,050
classification or pose estimation task which you can perform using Yolo V8.

114
00:09:13,980 --> 00:09:17,130
Uh, but you cannot perform this task with YOLO v9.

115
00:09:17,130 --> 00:09:18,180
With yolo v9.

116
00:09:18,180 --> 00:09:20,400
Currently, you can only perform object detection.

117
00:09:20,400 --> 00:09:25,230
You can train your object detection model, or you can train the object detection Yolo v9 object detection

118
00:09:25,230 --> 00:09:27,570
model on any custom data set as well.

119
00:09:29,260 --> 00:09:33,280
So here we have uh, yolo yolo v nine models.

120
00:09:33,280 --> 00:09:38,260
So YOLO v nine comes with four different models which are ordered by parameter code.

121
00:09:38,260 --> 00:09:40,510
You can skip the Yolo v 90 model.

122
00:09:40,510 --> 00:09:46,750
We will start from YOLO v9's, YOLO v nine medium, YOLO v nine compact and Yolo v nine extended.

123
00:09:46,750 --> 00:09:50,740
So YOLO v nine comes in four models orders by the parameter count.

124
00:09:50,740 --> 00:09:55,210
So now you can see that v nine small model has 7.1 million.

125
00:09:55,210 --> 00:09:55,810
So here I am.

126
00:09:55,810 --> 00:10:02,710
Represent millions 7.1 million parameters V nine medium has 20.0 million parameters.

127
00:10:02,710 --> 00:10:09,970
Yolo v nine compact has 25.3 million parameters and Yolo v nine extended has 57.3 million parameters.

128
00:10:10,270 --> 00:10:12,790
Yolo v nine comes in four models.

129
00:10:12,790 --> 00:10:18,370
Orders by the parameter count by nine small Yolo v nine medium, Yolo by nine compact, and Yolo v nine

130
00:10:18,370 --> 00:10:19,120
extended.

131
00:10:19,150 --> 00:10:23,020
Each model differs in terms of parameter, count and performance.

132
00:10:23,020 --> 00:10:24,460
So now you can see here.

133
00:10:24,670 --> 00:10:26,560
Here we have all the parameter in millions.

134
00:10:26,560 --> 00:10:32,500
And here we have the performance mean average precision on the validation set of the Ms-coco data set.

135
00:10:32,500 --> 00:10:38,860
So you can see that yolo V9S has the least mean average precision as compared to other YOLO v nine model,

136
00:10:38,860 --> 00:10:44,560
which is 46.8%, and Yolo v nine extended has the highest mean average precision on the validation set

137
00:10:44,560 --> 00:10:50,980
of the Ms-coco data set, which is 55.6%, and it has also has the higher number of parameters, which

138
00:10:50,980 --> 00:10:55,600
is 57.3 million as compared to the other Yolo v nine models.

139
00:10:56,790 --> 00:10:59,790
When it has more number of flops as well.

140
00:11:00,090 --> 00:11:02,460
So YOLO benign coco benchmarks.

141
00:11:02,460 --> 00:11:06,870
So the diagram below like you can see the diagram over here.

142
00:11:07,770 --> 00:11:08,460
So.

143
00:11:09,360 --> 00:11:10,440
You can see this diagram.

144
00:11:10,440 --> 00:11:16,080
The diagram below illustrates how the Yolo v nine models achieve high accuracy on the Coco dataset,

145
00:11:16,080 --> 00:11:22,200
while utilizing few parameters showcasing their efficiency in balancing model complexity with performance.

146
00:11:22,350 --> 00:11:27,030
So now over here, you can see that, uh, here we have on the x axis we have the number of parameters.

147
00:11:27,030 --> 00:11:33,120
And on the y axis we have the mean average precision on the validation set of the mscoco data set.

148
00:11:33,120 --> 00:11:33,870
And the model.

149
00:11:33,870 --> 00:11:40,650
We have evaluated the Yolo v nine model with other YOLO models on the Mscoco data set, uh, on the

150
00:11:40,650 --> 00:11:45,270
Mscoco data set, which is the benchmark data set to evaluate the object detection models.

151
00:11:45,270 --> 00:11:51,660
So over here you can see that this is this green line, this um, uh, you can say maroon color show

152
00:11:51,660 --> 00:11:54,450
the performance of YOLO V nine and the Glen model.

153
00:11:54,450 --> 00:11:58,110
So YOLO v nine paper publishes two different release, two different models.

154
00:11:58,110 --> 00:12:01,170
One is Glen model and other is YOLO v nine model.

155
00:12:01,170 --> 00:12:02,430
And here this.

156
00:12:03,370 --> 00:12:07,180
Maroon color show the performance of YOLO V9 and the Glen model.

157
00:12:07,180 --> 00:12:09,100
So this is the YOLO v9 model.

158
00:12:09,100 --> 00:12:11,140
And this is the, uh, Glen model.

159
00:12:11,500 --> 00:12:18,640
So you can see that YOLO v9 models uses less parameters and it outperforms in terms of accuracy than

160
00:12:18,640 --> 00:12:20,740
the previous Yolo YOLO model.

161
00:12:20,740 --> 00:12:23,470
So you can see that it outperforms Yolo V8 model.

162
00:12:23,470 --> 00:12:29,230
Yolo V6 model Yolov7 model Yolov5 model YOLO miss model gold YOLO model.

163
00:12:29,230 --> 00:12:37,030
So we can easily say that YOLO week nine uses less parameters and it like it is more accurate as compared

164
00:12:37,030 --> 00:12:39,550
to the other YOLO models YOLO models.

165
00:12:39,550 --> 00:12:45,790
So from the graph, we can say that YOLO v9 uses the less number of parameters, and it outperforms

166
00:12:45,790 --> 00:12:50,110
in terms of accuracy as compared to the other YOLO models.

167
00:12:50,110 --> 00:12:54,490
So YOLO v9 outperforms in terms of accuracy as compared to other YOLO models.

168
00:12:54,490 --> 00:12:55,330
So.

169
00:12:56,560 --> 00:12:58,240
Is the smallest model you v9's.

170
00:12:58,240 --> 00:13:00,070
So you can see over here the smaller.

171
00:13:00,280 --> 00:13:07,480
Smallest model YOLO V9's achieves 46.8 average precision on the validation set of the Ms. Coco dataset,

172
00:13:07,480 --> 00:13:16,150
while the YOLO largest model, YOLO v9 extended achieves 50.6 55.6% average precision on the validation

173
00:13:16,150 --> 00:13:18,370
set of the Ms. Coco data set.

174
00:13:18,370 --> 00:13:19,720
From here you can see.

175
00:13:21,640 --> 00:13:27,760
So now we'll do the comparison of YOLO v nine with YOLO v eight and yolov7 in terms of parameters count

176
00:13:27,760 --> 00:13:32,020
and flops, as well as in terms of mean average precision as well.

177
00:13:32,050 --> 00:13:37,960
So a quick overview of Yolo V8 Yolo V8 became popular because it offers a balance between speed and

178
00:13:37,960 --> 00:13:38,680
accuracy.

179
00:13:38,680 --> 00:13:45,130
So uh, basically YOLO B nine provides faster inference, faster inference with good accuracy, and

180
00:13:45,130 --> 00:13:50,860
it offers good real time performance, which makes it suitable for application, requires low latency.

181
00:13:50,860 --> 00:13:56,020
So inference means, uh, how much quick the output detection are.

182
00:13:56,020 --> 00:14:00,460
So you love it provides quick output detections with good accuracy.

183
00:14:00,460 --> 00:14:03,700
So which make is good for real time applications.

184
00:14:03,850 --> 00:14:05,350
Uh which requires low latency.

185
00:14:05,350 --> 00:14:06,940
So what is latency?

186
00:14:06,940 --> 00:14:11,290
In computer vision, latency refers to the delay or the time lag between the input.

187
00:14:11,290 --> 00:14:16,960
When we pass an image to the object detection model and the output where we get the image with bounding

188
00:14:16,960 --> 00:14:19,780
boxes coordinate, uh, around each of the detected objects.

189
00:14:19,780 --> 00:14:26,830
So the latency, uh, is the basically a time lag or that, uh, delay between the input when we pass

190
00:14:26,830 --> 00:14:29,470
the input to the object detection model and the output.

191
00:14:29,470 --> 00:14:36,310
So the time it takes, uh, for the process, uh, during the processing when we pass an input and we

192
00:14:36,310 --> 00:14:37,180
get the output.

193
00:14:37,180 --> 00:14:40,420
So that time is basically refers as latency.

194
00:14:41,480 --> 00:14:44,330
So in Europe, hate has very low latency.

195
00:14:44,330 --> 00:14:50,600
And in my experiments, which I performed, Yellow Gate has low inference or lower inference speed or

196
00:14:50,600 --> 00:14:53,300
latency as compared to yellow line.

197
00:14:53,300 --> 00:14:56,840
So this these are from my experiments in yellow eight.

198
00:14:56,840 --> 00:15:02,240
Low latency means that yellow eight can process images quickly, quickly, and provide results in real

199
00:15:02,240 --> 00:15:08,180
time or near real time, which makes Yolov2 suitable for application where quick responses are required,

200
00:15:08,180 --> 00:15:12,800
such as autonomous driving, real time anomaly detection or surveillance systems.

201
00:15:12,800 --> 00:15:17,210
So this is the comparison of Yolov7 Yolo v eight and Yolo v nine model.

202
00:15:18,380 --> 00:15:20,750
So in terms of parameter count the following.

203
00:15:20,750 --> 00:15:25,820
So these are the comparison of yolov7 yolo v eight and YOLO nine best models.

204
00:15:25,820 --> 00:15:28,730
So the Yolo v eight best model is Yolo V8X model.

205
00:15:28,730 --> 00:15:35,180
Yolo v nine best model is YOLO v nine extended model, and Yolov7 best model is yolov7 x model.

206
00:15:35,180 --> 00:15:39,290
So these are the parameters count number of model parameters in millions.

207
00:15:39,290 --> 00:15:42,080
So these are the millions in millions parameter count.

208
00:15:42,410 --> 00:15:51,200
So Yolov7 best model has 71.3 million parameters, while Yolo v eight best model has 68.2 million parameters,

209
00:15:51,200 --> 00:15:57,860
while Yolo v nine extended, which is the best YOLO v nine model has 58.1 million parameters.

210
00:15:57,980 --> 00:16:06,050
So we can say that in terms of parameter count count, there is a 15% decrease as compared to YOLO eight

211
00:16:06,050 --> 00:16:12,290
in terms of in the parameter count of Yolo v nine, while if we do the comparison of Yolo v nine with

212
00:16:12,290 --> 00:16:19,910
yolov7, there is a 19% decrease in terms of parameter count as compared to yolov7.

213
00:16:21,030 --> 00:16:24,750
And if we see the number of floating point operations in gigaflops.

214
00:16:24,750 --> 00:16:33,570
So yolov7 x 180 9.9, while in Yolo V8X, the number of floating point operations in gigaflops increased

215
00:16:33,570 --> 00:16:35,460
to 250 7.8.

216
00:16:35,550 --> 00:16:40,380
And in Yolo V9E it decreases to 190 2.5.

217
00:16:40,380 --> 00:16:48,600
So if we do the giga floating point uh operations comparison between Yolo v nine and Yolov7.

218
00:16:48,600 --> 00:16:55,920
So we can say that Yolo v nine has a bit, uh, more than uh, floating point operations as compared

219
00:16:55,920 --> 00:17:01,080
to Yolov7, but it has less floating point operations as compared to Yolo v eight.

220
00:17:04,550 --> 00:17:10,070
In terms of average precision, YOLO v9 outperforms other object detection models.

221
00:17:10,250 --> 00:17:19,640
YOLO v9 achieves an average precision of 56 55.6% on a validation set of the Ms-coco dataset.

222
00:17:20,060 --> 00:17:21,140
So we can, uh.

223
00:17:21,140 --> 00:17:24,470
It also outperforms YOLO v eight as well as Yolov7.

224
00:17:24,470 --> 00:17:32,240
With YOLO v eight, we get an average precision of 53.9% on the validation set of Ms-coco data set,

225
00:17:32,240 --> 00:17:35,960
and with Yolov7 we get only 52.9%.

226
00:17:35,960 --> 00:17:43,640
So we can say that Yolo v nine outperforms YOLO v eight and Yolov7 in terms of average precision on

227
00:17:43,640 --> 00:17:49,430
validation set of the Ms-coco data set, which is a benchmark data set to evaluate object detection

228
00:17:49,430 --> 00:17:50,060
model.

229
00:17:50,180 --> 00:17:56,330
So in conclusion, we can say that Yolo v nine achieves high accuracy and speed in object detection

230
00:17:56,330 --> 00:17:59,180
while reducing model complexity as you can.

231
00:17:59,180 --> 00:18:06,590
We have seen that yolo v nine uh gives good accuracy with less number of parameters and computational

232
00:18:06,590 --> 00:18:07,130
demands.

233
00:18:07,130 --> 00:18:13,700
This is evaluated by its performance on the Coco data set, where it demonstrates improvement with fewer

234
00:18:13,700 --> 00:18:19,070
parameters and less computational overhead compared to the other versions of the YOLO models.

235
00:18:20,410 --> 00:18:26,110
Thus you do not nurse with its unique architecture, using less parameters and less calculation flops

236
00:18:26,110 --> 00:18:29,410
and giving significant improvements in performance.

237
00:18:29,410 --> 00:18:37,060
So YOLO v9 has its unique architecture, in which we have integrated PGI and Glan so that we can, uh,

238
00:18:37,900 --> 00:18:39,910
prevent from the information loss.

239
00:18:39,910 --> 00:18:44,920
Plus we can also we have also seen that YOLO benign uh, outperforms in terms of accuracy.

240
00:18:44,920 --> 00:18:50,860
And it also uses less number of parameters as compared to the other state of the art YOLO models.

241
00:18:50,860 --> 00:18:59,050
So YOLO v9 outperforms other YOLO models by a very margin, uh, due to its unique architecture, and

242
00:18:59,050 --> 00:19:04,540
it uses less number of parameters and it gives better accuracy as compared to the other YOLO models.

243
00:19:04,540 --> 00:19:06,100
So that's all from this tutorial.

244
00:19:06,100 --> 00:19:07,150
Thank you for watching.