My name is Polong Lin, and I'm a data scientist/data science advocate, I guess. Okay. So just a quick overview of Watson versus OpenCV, before I jump into each of them in detail. Watson as you may have heard of it, is actually not a single entity, it's not a super computer, it's not just one thing, it's actually collection of things. So we've also made Watson available to everyone through a set of Watson APIs. So one of those APIs is called Watson Visual Recognition which allows you to do facial detection, some facial recognition, some object classification, all through a set of APIs, and it's very well documented as well. It comes with these pre-packaged models which makes it very easy for anyone to get started with. You create custom classifiers as well. So if you want to train your own models, all you have to do is have your own images uploaded to Watson, have it figured out for you, and it can train that custom classifier for you, and you can test it out yourself. So we will be trying that out today as well. Watson Visual Recognition because it's an API service, I wouldn't quite recommend it for real-time videos, or real-time facial detection, or recognition, but otherwise, it's certainly scaleable and of course, there is some cost after a certain free trial limits as well. OpenCV on the other hand, is an open source of course Python package. It comes with a set number of prepackaged models that allow you to do facial facial detection, and you even run facial recognition as well. It can also create custom classifiers like Watson, but it's quite a bit more difficult. In fact, as we could probably all tell from just the beginning of the session, installing OpenCV is a difficulty in itself. With OpenCV locally, you can certainly handle real-time videos as long as you can figure how to process and speed up your code. But there is a learning curve. All right. So at OpenCV there's a number of tutorials out there, but it can be a little bit tricky to learn. So if you really interested in jumping right into visual recognition, object classification, Watson is a great start. If you want to follow that up with some more customizability, then maybe you can look into OpenCV. Of course, OpenCV because of its open source nature, there is some bugs, but it's free. So let's begin with Watson. Watson as I mentioned, as many of you have probably heard of, one this can have a Jeopardy contest against Ken Jennings and Brad, I don't remember his last name, but Watson without using any connection to the Internet was able to answer a lot of questions. So this generated a lot of hype around Watson and perhaps give the impression that Watson was this one thing, that's one entity, and it really isn't. It's collection of things. It's applied in many different industries from healthcare, to banking, to insurance. So some of these capabilities Watson had been made available to everyone through these Watson API services. You can create chatbots with Watson. You can you can use APIs for text to speech or speech to texts. You can say something to Watson and Watson can transcribe that for you or the reverse. It can help you translate some text, determine the personality of a text as well. All these ones, the one that will focus on tonight is Visual Recognition. So this is what Visual Recognition on Watson looks like. This is the tool, this is the GUI. So this is the easy way to use the Watson Visual Recognition, which I think is a great way to get started, just to understand how Watson Visual Recognition works. So here, we have this tool. This is what you would start off with if you created your own free Instance. You have a default classifier, you have your food classifier, and you have your face detection classifier. All right. These are all pre-trained. They've been trained on millions of images already, on both positive and negative images. So these are pre-trained. If you give it an image, you simply drag and drop an image into Watson. So here, I found an image of a tank trying to challenge Watson and see if you could identify properly and it determines that it's an armored vehicle, to vehicle panzer or a tank, ultramarine color, army tank, and also shows you it's confidence scores as well. So that's pretty neat. That's why I generate object classification thing. For food, so you can upload images of sushi, of pasta, of cake, and it can also identify that without giving any labels prior. So it's able to identify this is a California roll or a salmon roll with some level of confidence. With face detection, so here I've uploaded, it's just less than a month away I think from the Game of Thrones Season 7 Premier. So here, I have uploaded the cast of Game of Thrones and detects not only their faces, but also the gender, and also the age range that Watson thinks that face belongs in. Then there's also this little button at the very top called Create Classifier, which allows you to create your own custom classifiers. So with Create Classifier, it's pretty simple, the concept. What you do is you can upload images, these images can be in zip files for each class that you want to create a custom classifier around. So let's say you have some images around apples, and you have images of bananas, and you have images of pears. So you can upload that to this custom classifier called fruits classifier. Once it can learn the different between apples, bananas, and pears. It can also learn the negative examples that you may want to upload it. So for example, vegetables, many of them, of course, look like fruits, or fruits look like vegetables, but you don't want vegetables to be classified as fruits. So you can upload images of vegetables so that when you end up submitting a vegetable, Watson will know that this is actually a vegetable and not a fruit. With Watson, you can also update your custom classifier as well. So if you have more images of apples, you can of course, update and add more images to that apple's class. So becomes smarter over time. If you have other classes, like let's say now we want to add oranges, we can add oranges as a separate class. Now, it can learn to identify oranges, whereas before, it would not have been able to do so. We can also add in cheese for example, as a negative example. So if someone uploads an image of a cheese, it would be identified as not a fruit. So this is very simple explanation of how the Visual Recognition Service works. First, at the very bottom and bottom left-hand corner, you see prepared training data. So these are your images of your apples, and oranges, and so on. Then you use that, you upload that to Watson and Watson will train and classify this custom classifier for you. Then in the middle, you bring in these images that you want to be classified. As long as Watson has trained sufficiently on your images, you will be able to see results, and hopefully, it will classify these images appropriately. So here's just a very quick example of something I did in Watson. So I had on my computer, 379 images of one of my colleagues Joseph. I actually collected this from OpenCV. You can collect images of faces however you'd like. You can even download Facebook images of your friends and you can use that as a classifier. It's part of a class about this classifier. So we have 379 images of Joseph, 250 images of myself, my face, and then of my manager Leon, 100 images as well. What I did was I took a zipped compressed file for each of these names, so Joseph, Polong and Leon. Then for each of these zip files, I uploaded them to Watson. All right. So this is under the Create Classifier section. So I can't draw it here, but you can see, I've created a class for Joseph and I've uploaded that zipfilejoseph.zip, for Leon, I've uploaded leon.zip, for Polong, I've uploaded polong.zip. In this case, I didn't upload any negative images. So I'm not looking at non Joseph, Leon, Polong faces. But I certainly could upload images of other people who are not Joseph, Leon, or Polong. Maybe celebrities, maybe images out there in the model. Then Watson creates that classifier. It can take a couple minutes. Then I upload these images that I've never used in the training dataset. So here's Leon there. That's his LinkedIn profile image that I had never used in the training dataset, and it determines that it's Leon with 71 percent accuracy. So this accuracy ranges between zero and one. If it's above 0.5, it's pretty good classification. I think that's as much as I'll say. Anything below that is not a very good confident score. So you can see Joseph and Polong at 0.13. For Joseph, it's at 0.36. At least, it's still quite a bit more confident that it's Joseph, and Polong, and Leon. Then for an image of me, it's at 0.34 and again, that's much higher than Leon or Joseph. So you can start to imagine, if you're a developer, you can take the first, the name of the highest confidence score, Leon, Joseph, and Polong, and use those as the classified labels for these images. This is quite extensive API reference for Watson Visual Recognition. It's available in both CRL, Node, Java, Python, lots of different SDKs which makes it very easy for anyone to get started.
