1
00:00:00,185 --> 00:00:01,785
So with that, good afternoon.

2
00:00:02,025 --> 00:00:03,945
It is my pleasure to introduce Dr.

3
00:00:03,945 --> 00:00:08,745
Vijay Kalivarapu, a Senior Artificial Intelligence Engineer at Pella Corporation.

4
00:00:08,745 --> 00:00:09,105
Dr.

5
00:00:09,105 --> 00:00:16,905
Kalivarapu holds a PhD in Mechanical Engineering with a co-major in Human Computer Interaction from Iowa State University.

6
00:00:16,985 --> 00:00:17,945
So big fan.

7
00:00:18,585 --> 00:00:25,785
He brings deep expertise in applying computer vision, machine learning, and 3D visualization to real-world manufacturing challenges.

8
00:00:26,065 --> 00:00:37,625
His work focuses on vision-based quality assessment, attribute detection, and precision measurement in production environments, where he has led multiple initiatives delivering scalable solutions with measurable business impact.

9
00:00:37,865 --> 00:00:50,985
In today's session, we will share practical insights from deploying vision AI in manufacturing, highlighting how these technologies improve quality assurance and measurement accuracy, as well as where they face limitations in real-world sessions.

10
00:00:50,985 --> 00:00:53,425
So please join me in welcoming Dr.

11
00:00:53,425 --> 00:00:54,345
Vijay Kalivarappu.

12
00:01:01,785 --> 00:01:09,225
So I kept wondering if I got the entire text of my presentation to Jake and have him read through a piece of paper.

13
00:01:11,065 --> 00:01:11,345
All right.

14
00:01:11,625 --> 00:01:12,505
Good afternoon, everybody.

15
00:01:12,505 --> 00:01:14,425
Welcome to my session.

16
00:01:14,785 --> 00:01:16,345
And thank you, Jake, for the introduction.

17
00:01:18,265 --> 00:01:26,185
So I'll go a little bit about our vision AI efforts in attribute detection and measurements.

18
00:01:26,705 --> 00:01:28,585
None of the work we did is

19
00:01:29,865 --> 00:01:44,585
groundbreaking cold fusion, but the effects of implementing them has been really good, and we are in a direction to implement more of those technologies in our company.

20
00:01:45,705 --> 00:01:47,065
Just a very quick overview.

21
00:01:48,905 --> 00:01:54,105
I've been with Iowa State for long enough to love experiments.

22
00:01:54,825 --> 00:01:58,425
I've been in the industry just enough to learn that

23
00:01:58,905 --> 00:02:01,705
I cannot take as much time as I want for experiments.

24
00:02:05,065 --> 00:02:12,025
So I do have background in 3D visualization, VR, AR simulations.

25
00:02:12,425 --> 00:02:14,665
I also teach design optimization at Iowa State.

26
00:02:14,665 --> 00:02:15,945
So it's a part-time gig.

27
00:02:16,905 --> 00:02:18,265
I have it at Iowa State.

28
00:02:19,065 --> 00:02:24,905
I am really excited to see that my semester is going to end in 10 days.

29
00:02:25,465 --> 00:02:25,705
So

30
00:02:26,225 --> 00:02:29,065
any faculty among you, would know the votes.

31
00:02:30,185 --> 00:02:32,505
So it's a little overview of my talk.

32
00:02:32,505 --> 00:02:38,425
I'm going to talk about who we are, what we do, and some of the general challenges we face at Pella Corporation.

33
00:02:38,745 --> 00:02:40,105
And I will go into the case studies.

34
00:02:40,105 --> 00:02:46,265
I'm going to talk about two of them, window attribute detection and screen door size violation.

35
00:02:46,265 --> 00:02:48,025
Well, not violation, validation.

36
00:02:48,505 --> 00:02:51,585
So a couple of things I will talk about each of those use cases.

37
00:02:51,585 --> 00:02:52,345
What is the objective?

38
00:02:52,385 --> 00:02:54,985
What is the existing or traditional practice

39
00:02:57,865 --> 00:03:04,425
With those two case studies, I will talk about how we implemented the Vision AI tech and then some of the challenges and outcomes.

40
00:03:04,985 --> 00:03:09,145
So first of all, Pella is a family-owned corporation company.

41
00:03:09,705 --> 00:03:12,105
Last year, we celebrated our 100th year.

42
00:03:13,545 --> 00:03:20,825
And one of the big push last year was to use modern tech to improve our manufacturing processes.

43
00:03:20,825 --> 00:03:22,505
And a big part of it is Vision AI.

44
00:03:22,985 --> 00:03:24,545
We have 11,000

45
00:03:25,145 --> 00:03:28,025
team members from across the country working for Pella.

46
00:03:28,225 --> 00:03:33,225
And then under the Pella umbrella, we have five different companies, and Pella is a flagship brand.

47
00:03:34,105 --> 00:03:38,385
So just a quick overview of where we are within the US.

48
00:03:38,385 --> 00:03:44,585
So Iowa has one, two, three, 4, 4 locations.

49
00:03:44,985 --> 00:03:53,385
All of them are wood plants, but then across the country we have plants that do aluminum, vinyl, and fiberglass as well.

50
00:03:54,825 --> 00:04:01,545
And most of the products that we make are designed for residential homes and commercial applications.

51
00:04:02,825 --> 00:04:07,065
So I want to talk a little bit about the general challenges.

52
00:04:08,185 --> 00:04:22,065
As opposed to a typical manufacturing company, like say an automotive company, ours is unique in that every part that comes to a station, workstation, is different.

53
00:04:23,385 --> 00:04:25,785
because our units are custom built.

54
00:04:26,185 --> 00:04:31,145
So that by itself produces various challenges.

55
00:04:31,785 --> 00:04:36,465
The second one is in wood plants, we also have other things to deal with.

56
00:04:36,905 --> 00:04:38,185
Wood defects, right?

57
00:04:38,745 --> 00:04:41,305
Cracks, knots, or pitch pockets.

58
00:04:41,545 --> 00:04:47,505
So those are not homogeneous and they don't appear at the same place in a lumber

59
00:04:48,025 --> 00:04:50,825
piece that goes through our factory floor.

60
00:04:51,225 --> 00:04:54,105
So how do we identify these defects?

61
00:04:54,465 --> 00:04:58,025
And can we use Vision AI to identify these defects?

62
00:04:58,505 --> 00:05:09,785
Some of the other issues, I mentioned this in the talk, well, in our discussion earlier in the main room.

63
00:05:10,265 --> 00:05:11,625
So how do we

64
00:05:13,225 --> 00:05:17,305
make the users of this tech trust in AI.

65
00:05:18,105 --> 00:05:19,705
It's not very easy.

66
00:05:20,105 --> 00:05:22,185
It's a generally an uphill task.

67
00:05:22,505 --> 00:05:29,305
So we as technology developers, we develop this tech.

68
00:05:29,625 --> 00:05:37,145
But if it fails at some point, then the end users are immediately going to throw it away, throw the towel.

69
00:05:37,385 --> 00:05:38,185
I'm not going to use that.

70
00:05:38,585 --> 00:05:41,385
So it's a big task.

71
00:05:42,025 --> 00:05:51,065
for us to develop this tech and also make sure the end users are okay with using the tech.

72
00:05:51,785 --> 00:05:56,905
The other things, vision AI opportunities, where do we use vision AI?

73
00:05:56,905 --> 00:06:02,185
This right from picking the material that builds a window frame.

74
00:06:02,665 --> 00:06:06,945
all the way down to whether the product is installed correctly or not.

75
00:06:07,305 --> 00:06:13,065
So each of these places have a spot for Vision AI being implemented.

76
00:06:13,465 --> 00:06:16,345
So I'm going to talk about two of those case studies.

77
00:06:16,665 --> 00:06:18,265
The first one is attribute detection.

78
00:06:19,305 --> 00:06:22,505
The objective, I just put it here, to the window units.

79
00:06:22,505 --> 00:06:26,585
I'm going to talk about window units, but it applies to doors as well.

80
00:06:28,825 --> 00:06:29,225
So

81
00:06:29,785 --> 00:06:37,065
Do the window units that went through the production, does it have all the attributes on it that are defined in the spec?

82
00:06:37,545 --> 00:06:40,585
So that's the general objective.

83
00:06:41,065 --> 00:06:46,345
But to get there, what is the traditional or existing process?

84
00:06:46,665 --> 00:06:48,585
It is basically checked manually.

85
00:06:51,145 --> 00:06:57,465
So people that work from station to station, they have to identify if there are any issues, if there are any missing parts.

86
00:06:58,905 --> 00:06:59,265
And

87
00:06:59,825 --> 00:07:05,625
and before sending it to the next station, they would have to make sure that everything is in there.

88
00:07:05,705 --> 00:07:06,745
But it is error prone.

89
00:07:07,225 --> 00:07:11,385
So can we use Vision AI to ease their burden?

90
00:07:11,865 --> 00:07:15,705
So in order to do that, we designed it as a two-step process.

91
00:07:16,025 --> 00:07:18,505
The first one is an image quality check.

92
00:07:18,825 --> 00:07:24,105
So we use cameras to capture images of various stations.

93
00:07:24,745 --> 00:07:29,145
And we want to make sure that the images we capture are good enough

94
00:07:29,625 --> 00:07:34,585
to determine whether all the parts that are supposed to be on the window unit are there or not.

95
00:07:35,225 --> 00:07:38,905
So image quality check followed by attribute check.

96
00:07:39,225 --> 00:07:50,825
So if you look at this picture, right, we want to filter out those images that don't have attributes that we prefer.

97
00:07:51,065 --> 00:07:52,905
So the first one, unit out of view.

98
00:07:53,625 --> 00:07:57,145
We want to have a window unit in there when an image is captured.

99
00:07:57,625 --> 00:08:11,305
So the way we have it is at a station, an operator, when he thinks that the product is ready to move to the next station, he clicks a button for the camera to capture an image.

100
00:08:11,545 --> 00:08:14,425
But sometimes he captures an image without anything in there.

101
00:08:14,545 --> 00:08:15,305
And that is an issue.

102
00:08:15,705 --> 00:08:17,545
And we would have to filter those out.

103
00:08:18,465 --> 00:08:28,905
And the second one, tilt table up, so that the table where they perform their work, it will be tilted up so that the heavy window unit can be transferred over from one station to another.

104
00:08:29,305 --> 00:08:37,785
But if we have the tilt table up, we don't see all the attributes in the window unit, so we'd have to somehow let the user know that, hey, you need to take a picture again.

105
00:08:38,465 --> 00:08:41,225
The third one, material obstruction, right?

106
00:08:41,225 --> 00:08:46,425
If there's a pair of gloves or any tools on a window unit that is obscuring

107
00:08:47,465 --> 00:08:50,985
some of the attributes that we want to identify in a window unit.

108
00:08:51,305 --> 00:08:55,625
So these are all under simple classification models.

109
00:08:55,705 --> 00:09:00,345
We can say there is a pair of gloves, there are tools, or there's a tilt table up.

110
00:09:00,745 --> 00:09:03,945
But it becomes complicated really quick.

111
00:09:04,745 --> 00:09:05,545
Take a look at this one.

112
00:09:06,585 --> 00:09:10,505
The left-hand side image, we have a cardboard person and tools.

113
00:09:11,305 --> 00:09:14,545
All three of those are obscuring the window unit.

114
00:09:15,945 --> 00:09:18,425
for an image to be captured properly.

115
00:09:19,225 --> 00:09:31,465
Now, the issue is if those 3 obstruction, pieces of obstruction, are not on a window unit, it is okay.

116
00:09:32,185 --> 00:09:36,345
But because it is obscuring the window unit, that would cause an issue.

117
00:09:36,585 --> 00:09:39,225
So classification models will not work for us anymore.

118
00:09:39,545 --> 00:09:41,145
So we ended up with using

119
00:09:41,945 --> 00:09:42,745
object detection.

120
00:09:42,985 --> 00:09:53,305
So that we can have our own logic to say if a window unit is fully exposed, it doesn't matter where the other things are.

121
00:09:53,545 --> 00:10:02,825
So that helped out with filtering out items that can go to the next step, which is our attribution detection.

122
00:10:05,225 --> 00:10:08,105
So assuming that we got a good image,

123
00:10:08,665 --> 00:10:14,025
We run it through segmentation models to identify different aspects of a window unit.

124
00:10:16,345 --> 00:10:25,225
We position our cameras in such a way that we get a good vantage point of a window.

125
00:10:25,225 --> 00:10:32,265
So it's not exactly top-down because we want some attributes detected around the edges of a window unit.

126
00:10:32,665 --> 00:10:36,345
So with that, we train some models.

127
00:10:37,465 --> 00:10:59,705
identify different aspects of different attributes within the model, and then check them against the spec of the window unit to tell there are parts missing in the window unit, or if there is an issue with the operator himself, if he did not, if he has incorrect attributes installed on the window unit.

128
00:10:59,705 --> 00:11:04,105
So those are the kind of things that we were able to implement using Vision AI.

129
00:11:04,185 --> 00:11:06,265
And this has worked out really well.

130
00:11:08,105 --> 00:11:15,705
With the models that we developed, we got to about 90 to 95% consistency, but we are shooting for more.

131
00:11:17,065 --> 00:11:20,185
And the other thing is we want these to be done at near real time.

132
00:11:20,585 --> 00:11:36,905
Because once the window unit is ready to go to the next station, there's not as much time, because the operator will not wait for 30 seconds before the inference is done and say we are missing a part or something.

133
00:11:37,385 --> 00:11:43,225
So we started using GPU-based inferences, and it worked out really well.

134
00:11:43,385 --> 00:11:52,265
So far, we have implemented them in Pella and one or two other facilities across Iowa, but we're trying to roll it out across the entire country as well.

135
00:11:53,625 --> 00:11:58,585
And the last thing is train the usage to shop floor and quality techs.

136
00:11:59,385 --> 00:12:01,785
So every once in a while,

137
00:12:02,665 --> 00:12:04,185
there are changes to our process.

138
00:12:04,345 --> 00:12:14,985
Like if we make the Pella logo sticker slightly different, that means that we are introducing error into inferencing our model, right?

139
00:12:15,145 --> 00:12:25,625
Or if we change the way a hardware pack is installed on the window unit, if that changes, it is going to add another line of error.

140
00:12:25,865 --> 00:12:27,945
So over time, there is a model drift.

141
00:12:28,185 --> 00:12:46,105
So what we do is we not only develop this tech, but we also write documentation procedures to teach quality techs to train models every couple months and then reintroduce those models back into our total workflow.

142
00:12:46,585 --> 00:12:48,665
So far, it has been working really well.

143
00:12:49,625 --> 00:12:52,905
We are implementing it on a larger scale now.

144
00:12:53,665 --> 00:13:00,185
So I'm going to switch over to the next case study, screen door size validation.

145
00:13:01,465 --> 00:13:14,505
The objective, it's not so much whether a screen door has been assembled right or not, but it's more of whether they are putting the right product into the right box.

146
00:13:16,345 --> 00:13:21,385
Because we've seen a lot of times when a screen door that is supposed to be 53 inches,

147
00:13:22,745 --> 00:13:25,945
But by the time it goes into the box, it is 64.

148
00:13:27,065 --> 00:13:35,545
So there are issues, manual issues, or human errors before screen doors are packed into a shipping box.

149
00:13:36,025 --> 00:13:47,545
So the question was, can we do a last check to tell, to give a rough dimensions, or to get rough dimensions of a screen door that is being put into a box?

150
00:13:48,265 --> 00:13:52,345
So we applied Vision AI, and it's a three-step process this time.

151
00:13:52,505 --> 00:13:54,425
The first one is an image quality check.

152
00:13:55,465 --> 00:13:59,385
So we filter out people that appear in the image.

153
00:13:59,705 --> 00:14:03,625
The next one is to overcome camera lens distortions.

154
00:14:03,985 --> 00:14:04,745
What does this mean?

155
00:14:06,905 --> 00:14:16,425
So typically, every camera that we use, the lenses that come with them, they have certain optical defects.

156
00:14:17,385 --> 00:14:23,945
and it cannot be avoided, but we can use certain pieces of information to correct it.

157
00:14:24,425 --> 00:14:35,785
So the three main ones that these cameras come with are barrel distortion, where images sort of appear rounded, or sometimes they have thin cushion effect.

158
00:14:36,265 --> 00:14:38,105
So if you take a picture of...

159
00:14:39,265 --> 00:14:49,545
of, say, a pillar from a distance using your phone, you would probably realize that there is some curviness to the pictures you've taken.

160
00:14:49,545 --> 00:14:51,305
And that's because of the lens distortion.

161
00:14:51,865 --> 00:14:55,625
So we have to overcome this distortion in order to do any kind of measurements.

162
00:14:57,225 --> 00:15:03,225
And then apply computer vision methods to tell what is the length of a certain part.

163
00:15:03,705 --> 00:15:06,905
So overall, we filter out

164
00:15:07,705 --> 00:15:09,065
people using object detection.

165
00:15:10,425 --> 00:15:15,785
And this is, say, an input image or an image capture from a camera.

166
00:15:16,265 --> 00:15:18,425
So we do calibration process.

167
00:15:19,785 --> 00:15:23,065
The cameras that we use, they are pretty dumb cameras.

168
00:15:23,145 --> 00:15:28,185
They just take pictures, they stream images or videos, but they don't do anything else.

169
00:15:28,665 --> 00:15:35,545
And it also gives us more opportunity to take control because we can program it in a way that we wish.

170
00:15:36,185 --> 00:15:47,505
So, what we do is we calibrate the cameras so that we know the focal length in X direction, Y direction, what is the optical center of those lenses, and what are...

171
00:15:47,585 --> 00:15:49,065
the distortion parameters.

172
00:15:49,065 --> 00:15:58,025
So these are very standard computer vision methods of approaches to calibrate a camera.

173
00:15:58,345 --> 00:16:03,105
So once we calibrate the camera, we can do something called as undistortion.

174
00:16:03,465 --> 00:16:14,185
Now we can see that the image, the edges of the table are straightened out, which means that I have something to work with, right?

175
00:16:14,185 --> 00:16:15,665
So I can do some measurements.

176
00:16:16,105 --> 00:16:28,825
But in addition to undistortion, I can also do perspective correction where I change the, transform the image so that I get an output image, something very similar to this.

177
00:16:29,465 --> 00:16:39,305
Meaning that I can transform the image as if the camera is facing top down on the window unit itself.

178
00:16:39,785 --> 00:16:45,225
And because we know the dimensions of the tilt table, I can use that to map

179
00:16:46,145 --> 00:16:48,585
on to what the size of a screen door is.

180
00:16:49,145 --> 00:16:56,345
So using that, we were able to get to about a tolerance of less than 1/4 of an inch.

181
00:16:56,665 --> 00:17:04,265
So within 1/4 of an inch, we were able to identify the dimensions of a screen door, and it has helped us tremendously.

182
00:17:05,065 --> 00:17:07,945
So 2 cases talked about.

183
00:17:10,385 --> 00:17:12,425
Vision AI is not,

184
00:17:14,825 --> 00:17:17,585
It's not something that is really hard to implement.

185
00:17:17,785 --> 00:17:34,865
You just have to put those pieces together, see what works with building a machine learning model, and use that information to perform computer vision operations and make the tech help us.

186
00:17:34,865 --> 00:17:41,305
So that's where we are at, and we've been working with various other initiatives as well.

187
00:17:42,025 --> 00:17:42,665
That brings us

188
00:17:43,465 --> 00:17:44,905
to the end of my talk.

189
00:17:45,425 --> 00:17:47,465
If you have any questions, I'll be glad to answer.

190
00:17:51,345 --> 00:17:52,505
I have a question.

191
00:17:52,505 --> 00:17:53,065
Real quick.

192
00:17:53,785 --> 00:17:55,465
Thanks for sharing your case studies.

193
00:17:55,945 --> 00:18:01,225
Can you give us some insight into how long this took to maybe pull together?

194
00:18:01,225 --> 00:18:03,185
Did you do it internal with your team?

195
00:18:03,185 --> 00:18:04,265
What's your team size?

196
00:18:04,265 --> 00:18:05,465
Did you have outside help?

197
00:18:05,465 --> 00:18:09,705
Can you give a little insight just in case any of us want to do something like this?

198
00:18:10,505 --> 00:18:13,225
I will try to give as much info as I can.

199
00:18:13,225 --> 00:18:16,505
Chris, please correct me if I did anything wrong.

200
00:18:17,945 --> 00:18:21,545
So the Vision AI team is small.

201
00:18:21,945 --> 00:18:25,625
We have four to five people on our team.

202
00:18:26,825 --> 00:18:31,545
But we use an external tool to label our images and build models.

203
00:18:34,345 --> 00:18:38,105
The biggest challenge with, not challenges, the biggest time taker is

204
00:18:39,225 --> 00:18:40,665
is labeling these images.

205
00:18:41,305 --> 00:18:48,265
So we first have to identify what are the classes that we want to figure out in an image.

206
00:18:48,585 --> 00:18:54,985
So on a screen door, what I wanted to know is within the image, where is the screen door?

207
00:18:56,025 --> 00:19:00,905
So we have a collection of different screen doors with different colors, different sizes.

208
00:19:02,345 --> 00:19:07,225
So just like how a kid learns how to differentiate between an apple and an orange.

209
00:19:07,465 --> 00:19:22,585
He sees apples and oranges in different environments so that even at the end, if I take an apple, put it in the corner of a room, and ask him to tell what it is, to be able to tell that it is an apple and not an orange.

210
00:19:22,825 --> 00:19:28,185
So basically train the model with a whole bunch of images of a screen door.

211
00:19:28,345 --> 00:19:30,345
Same thing with our other attributes as well.

212
00:19:30,665 --> 00:19:32,585
So we train the model,

213
00:19:34,665 --> 00:19:42,745
Depending on whether we want to use it for image quality, that means object detection or attribute detection.

214
00:19:43,305 --> 00:19:53,225
So if it is object detection, just to tell me where roughly a tilt table is, a rectangle would be enough.

215
00:19:54,025 --> 00:19:57,465
But if I wanted a specific attribute of a specific size,

216
00:19:58,265 --> 00:20:00,345
then we would do a polygonal labeling.

217
00:20:00,585 --> 00:20:08,265
So that means we can have a customized polygon that represents a specific item.

218
00:20:08,825 --> 00:20:13,705
So these are the ones that take the longest.

219
00:20:14,505 --> 00:20:22,345
But once the model is built, we can run a quick script that takes in an image, spits out

220
00:20:23,905 --> 00:20:28,425
and tells us what attributes are identified or detected within that image.

221
00:20:28,985 --> 00:20:31,505
So is this like a few-month project, a few weeks?

222
00:20:32,225 --> 00:20:35,105
Or what's your time table to have this finished product?

223
00:20:35,465 --> 00:20:49,705
So if we have the images labeled correctly, the model training time itself takes anywhere from a couple of hours to a couple of days.

224
00:20:51,225 --> 00:20:54,505
But in order to get there, are a lot of learnings, though.

225
00:20:54,945 --> 00:20:57,705
And that's the one that took us most time.

226
00:20:57,945 --> 00:21:04,265
So we started on this approach a year and a half to two years ago.

227
00:21:04,265 --> 00:21:06,185
I wasn't with Pella at the time.

228
00:21:06,265 --> 00:21:07,545
I only started last year.

229
00:21:08,905 --> 00:21:14,185
But by the time we started, we made some progress, and we did learn some things.

230
00:21:14,905 --> 00:21:20,905
Whether or not, whether it's object detection or segmentation models, that is something that we learned through the process.

231
00:21:22,065 --> 00:21:29,225
and what are the resolution of images that we need to punch into our model?

232
00:21:29,865 --> 00:21:40,985
Do we pick a 256 model image, or do we want to have a 1024, 1024, or do we want to go all the way to a 2K or a 4K image?

233
00:21:41,305 --> 00:21:45,145
So all of that took a little bit of learning.

234
00:21:45,545 --> 00:21:48,985
So it's a subjective answer.

235
00:21:51,865 --> 00:22:00,505
If attributes are well-defined and there's the same ones, then you don't need as much time.

236
00:22:00,985 --> 00:22:05,065
But when objects do change, yes, it does take time.

237
00:22:05,305 --> 00:22:07,305
We have time for one more brief question.

238
00:22:07,745 --> 00:22:09,305
And I think I saw a hand back here first.

239
00:22:09,305 --> 00:22:10,265
Do you want me to go to him?

240
00:22:10,265 --> 00:22:10,465
Okay.

241
00:22:11,065 --> 00:22:12,425
And then we'll transition speakers.

242
00:22:14,185 --> 00:22:17,985
Yeah, two very short questions.

243
00:22:17,985 --> 00:22:19,145
One, I was just curious,

244
00:22:19,505 --> 00:22:24,825
what the average data set size is for like a single model of like a screen door.

245
00:22:24,945 --> 00:22:30,265
Are you talking like 100 images or a few thousand images for a single model?

246
00:22:30,385 --> 00:22:37,625
And the other is, the data just logged for later reference or is if there's a mistake, does a light come on?

247
00:22:38,025 --> 00:22:41,225
Yeah, for the operator to say, oh, I put this together wrong.

248
00:22:41,705 --> 00:22:41,945
Yes.

249
00:22:42,345 --> 00:22:47,065
So the first one, your first question was.

250
00:22:47,145 --> 00:22:48,185
Size of the model.

251
00:22:48,745 --> 00:22:49,545
size of the model.

252
00:22:51,465 --> 00:22:57,145
For screen doors, we had a few hundreds of images, hundreds of images.

253
00:22:59,305 --> 00:23:00,865
They're pretty much homogeneous.

254
00:23:00,865 --> 00:23:02,345
They're pretty much consistent.

255
00:23:02,905 --> 00:23:10,425
But for attribute detection, we're talking about thousands, like 4 or 5 thousands, or sometimes even more.

256
00:23:12,425 --> 00:23:13,865
So the other question you had was...

257
00:23:13,945 --> 00:23:18,105
Is the data just logged for later reference?

258
00:23:19,225 --> 00:23:25,705
Or is there something to help an operator, the person that built the screen door that they did something wrong?

259
00:23:25,945 --> 00:23:26,265
Yes.

260
00:23:26,425 --> 00:23:30,185
So we have a heads-up display, a TV monitor basically.

261
00:23:30,425 --> 00:23:41,545
So after an inference is done, if it detects that some of the attributes are missing, then it does, we send the inference picture

262
00:23:42,985 --> 00:23:45,865
to the heads-up display to tell that, hey, something is wrong.

263
00:23:46,105 --> 00:23:48,505
So it can go green, orange, or red.

264
00:23:48,665 --> 00:23:52,665
Red means you have to stop and you have to address it.

265
00:23:52,905 --> 00:23:54,025
Orange, it is okay.

266
00:23:56,505 --> 00:24:01,145
All right, folks, that's our time for this one, but I'm sure Vijay will be around afterwards if we have more questions.

267
00:24:01,145 --> 00:24:02,105
So thank you.

268
00:24:02,665 --> 00:24:05,545
Everybody join me in thanking Vijay for his wonderful presentation.

269
00:24:09,585 --> 00:24:09,705
And