When working with image classification, you cannot have a one size fits all approach. Some models work well at identifying a certain set of images, while may not work for another set of images. IBM Watson visual recognition has custom classifiers that allow us to specialize the type of images we want to classify. You might also wonder what would happen if you upload a tricky images to Watson. In this example, we have the famous drawing of the "Duck-Rabbit" which depending on how you look at it, looks like either a duck with its bill facing left or a rabbit facing the right with its ears behind its head. If you send this image to IBM Watson Visual Recognition, you might hope to see duck or rabbit but it actually struggles. The general classifier labels ash Grey color, animal, and mammal but then it gets some animal names wrong thinking that it might be an otter, sloth or rodent. So why does this happen? Before I answer this, let's try another tricky example. What about this image of fishing lures? We can tell they're lures because of the hooks though they were made to look like actual fish. So does this fool IBM Watson? Well, if you send this image to the general classifier, IBM Watson certainly does seem to get fooled. It identifies fish, aquatic vertebrate, animal, and so on. But it makes no mention of fishing lure. Yet, if you send the IBM Watson this image, now it is able to identify bait, fisherman's lure, and fishing lure. So what's happening? Why isn't computer vision perfect at classifying images? The primary reason is that it depends on how the classifier was trained. Let's say you grew up seeing various cats and they all seem to have some things in common. They are furry four-legged animals with triangular ears, a tail, sharp teeth claws, and whiskers, wonderful and cute animals and then one day your friend invites you over to their home and this is what you see. It has four legs, pointy ears, claws but no fur and no whiskers but it's still a cat. Actually, a breed called the Sphynx cat. So why was it more difficult to immediately recognize it as a cat? Well, that's because when your human brain was learning what cats were by seeing various examples of cats, it developed this model of what a cat should look like. So when it saw a Sphynx cat, it had troubled generalizing from the traditional understanding of furry cats to a completely hairless cat. Similarly, an image classifier, such as IBM Watson, will also run into difficulties if it hasn't been trained with enough examples of a particular class of objects or if it fails to generalize to something that's quite different from what it's used to seeing. On the other hand, it could be that the objects are simply really, really hard to distinguish between other objects. I mean look at these images. Have you ever doubted that you would have an easy time distinguishing between a dog or a mop or that distinguishing between Chihuahuas and muffins would be more difficult than you'd expect? Some objects could just happen to share a lot of similar features as other completely irrelevant objects and that might make it difficult for typical computer vision models to do a good job at classification. So now that we know the benefits and limitations of computer vision with IBM Watson, let's recap what we can do with the IBM Watson visual recognition service. With IBM Watson visual recognition, you can choose between which classifiers you want to use to classify images and there are a number of these classifiers like the general classifier, faces and food classifier. They're called Pre-trained because the IBM Watson team has already conducted the mundane part of training the classifiers with millions of images, allowing you to simply upload your own images to get a response of labels and so if you're interested in identifying objects across a wide breadth of categories, then the general classifier might make more sense to you. If you are primarily interested in identifying foods or to detect explicit images, then you can use those respective classifiers. In fact, there are other classifiers available that you will be exploring in the lab exercise. Most importantly, the classifier you choose to use should be dependent on the purpose or application that it's serving. If you're going to create an app that helps the blind use their phone cameras to identify food at the grocery store, then you'll probably want to use an image classifier that specializes in food and you probably wouldn't care whether or not the people show up in the camera. You can also create custom classifiers to that don't exist with IBM Watson. For example, to classify between different models of cars or years manufacturer, fresh food versus spoiled food and so on. But hold that thought on custom classifiers. For now, we'll be covering how to create our own custom classifiers later on. Thank you for watching this video.
