Name: 可口可樂公司使用TensorFlow進行數字營銷活動（TensorFlow遇見）。 (The Coca-Cola Company using TensorFlow for digital marketing campaigns (TensorFlow Meets))
Uploaded: 2021-01-14T10:48:38.000Z
Duration: 7 min 46 s
Description: 【看影片學英語】數萬部 YouTube 影片，搭配英漢字典即點即查，輕鬆掌握單字發音與用法，長久累積看電影不必再看字幕。

And welcome to the Tensorflow Cafe at the TENSORFLOW Developers Summit.

I'm Laurence Maroney, and I'm chatting with Patrick Brand from the Coca Cola Company on Patrick.

So I shared how tensorflow supporting one of our largest, most popular digital marketing programs here in the United States.

Okay, so it's like it's scanning codes from bottle caps.

So we've, uh, underneath a lot of the bottle caps in fridge packs, the products we sell.

We've got unique codes that are consumers can then use to intern to promotions and earn and win things.

Um Well, the we started, actually, early on on, I could recognize about last year, actually, and I could recognize that tensorflow had a very I call it approachable interface.

You know, it presents what would be otherwise very complex constructions in idioms that I think I would say every day program is gonna understand.

So you've got modules, got objects, you got classes.

You can compose these things to create, like a deep learning graph.

So it was that it was the entire, you know, platform.

The documentation, everything it was it seemed like something that was Ah, you know, a good move for us.

So you trained it to classify basically the digits that are printed on his bottle tops.

And now I've seen these bottle tops, and sometimes I'm sure that dot matrix printer printing on these things, it's like, sometimes their skin, and sometimes they're obfuscated.

So that was one issue we had had with our caps on other codes.

They couldn't handle the fidelity of the caps.

So what we did is we created a training set, both synthetic and real world, because we can generate these things these caps and fridge packs ourselves and really just, you know, through to tensorflow.

And, uh, we adjusted our model several times for development, but we eventually got something that performs extremely well for us and very well at scale.

Okay, so would you say it scaled so pretty quick?

Yeah, it's very fast, you know, it's ah, we're about one second.

Quiet one second average processing time from a user experience perspective is great.

You know, it's it's not ruining anyones expectation when they engage with our platform.

What's the difference if he had done this more and more traditional programming, Right?

So if we didn't say applied Amore heuristic model right, trying to use rules to interpret these codes so I could imagine a situation right, like where we create our model and it works really well for Let's zero degree rotation on the cap.

Users are holding this thing perfectly and holding their phone perfectly the second they twist that thing two degrees in the direction that heuristic model would fall apart.

So you find yourself having to in code all these different variations heuristics, And that sounds arduous, definitely, because the cap circular you don't really have a frame of reference.

I know that it's been yeah, I mean, just simple.

Something simple is a slight twist from the access and then rotation, which is very natural for someone holding a cap in their hand to do so.

You know, with a machine learning approach, Um, we're able Thio accommodate that translation appearance with convolution of neural net.

Sure, I think when the things I found that, like just to my eye that when I look at these digits on the bottle caps, it's sometimes hard to tell.

I have to tell the six from a five or a NATO that kind of thing.

So did you have a pretty good success with?

Absolutely, Um I can say anecdotally that Ah, that the machine I, uh, if you want, think about that way is actually often works better than the human eye just for the reasons you were saying that to him and I a lot of the, um, visual artifacts that are on these caps.

What's Q and pixel drift can start to complicate things, but our model is trained to look through that.

Okay, So how did you get training data, by the way?

Uh, where we started with synthetic data set.

We actually got I would say about 50% accuracy just on synthetic data alone.

Actually, so we created a, well, a real world data set, just printing stuff way dialed up our printing facilities and we said, Hey, give us a bunch.

And then, um, I just want the whole bunch of just a whole bunch of caps for giving.

The multiple suppliers, and we drive them with some custom built tools that we created toe.

Make it easy for them to take a picture and label it.

We started moving into a much more similar line approach.

One team would take images, and the other team to make a portal would verify the image entering the code.

Make sure the quality of the image itself was good.

But we were able to very quickly, within a matter of weeks, generate enough data for us to train the model.

Now, you mentioned that there are a number of initiatives you're working on, not just this one.

I can, um, so in particular with computer vision on marketing, this code scanning initiative was the beginning of our kind of our computer vision platform.

We this year have launched a new digital icon visual icon that we're starting to put on the exterior of our packaging.

And it's a delightful looking, very beautiful image, and it fits in well with all of our other branded elements, which is critical, but it encodes information for that product.

I would think of it more as like a bespoke Q R code, something that is appealing from a visual.

So instead of just lots of black and white squares, it's yes, it's three concentric rings with a little Coke bottle in.

Not all of our brands have these caps primitive, but but most of the Newman, certainly our core brands.

You know any Coca Cola product, our aid, et cetera.

Um, but you byproduct twist off cap, and you now have everything you need.

You know, you would create a count coke dot com slash rewards are on the new app that we launched two days ago.