Welcome to the third course in this specialization focuse on using advanced techniques and Tensor Flow. In this course, you lavish everything, you master it in the first two causes to solve some exciting problems in computer vision. Whereas there's a lot that you can do with sequential models in computer vision. These are models that take the data and process it one layer at a time policy forward. There are also a lot of more advanced things that cannot be handled with the standard sequential model. Take object detection. Your models like Yolo, the various flavors of Yolo or retina yet and wish you don't want to just take us into the picture and say, Is this a cat or not? But you may also want to draw a bounding box around the cat or around multiple cat or multiple objects that needs a more complex outputs in a more complex set of steps than the standard sequential model or take image segmentation models like unit where rather than just recognizing cats, you might also want to label every pixel as belonging to an object blind to the cat or not and this allows you to very accurately draw a boundary, we draw the outline of the cat in the picture. That also will require something beyond the standard sequential model. One of my favorite examples comes or class activation maps. I was doing work in diagnosing from X-ray images and it turns out that if you have a chest X-ray and a neaural network suggested diagnosis suggests, well, this patient has cardio megaliter and enlarged heart. A clinician might not want to just have your neural networks output. They may want to know what part of the X-ray image was the neural network looking at in order to render that diagnosis is quite reassuring. If you can have an algorithm such as a cam class activation, there that highlights the area of the heart in the chest X-ray to show the clinician that, yes, we looked at the heart and is looking at the heart there. We think this patient may have a case of an enlarged heart. In this course you'll learn about the APIs that allow you to go beyond the standard sequential model to build models like these, object detection or image segmentation or things like the camera. I'm delighted to have again Lawrence as instructor for this course, so please tell us more, Lawrence. Yeah, it's really great to be here. Thanks Andrew and computer vision is to me, my favorite part of working with deep learning and machine learning and AI. The possibilities that it opens, like he mentioned, with chest X-rays and being able to do diagnoses like that, I just find it really inspirational. In this course, one of the things that I really wanted to get to was we've learned previously about how to do things like convolutional neural networks using the sequential API. But can we take out much further to do more exotic models like you've mentioned Unet. But first, to be able to do that as a colder, we really need to be able to be much more lower level do things like with the functional API, because for example, with unets, you've got that lovely U-shape and your data flows down into the U, but then there's also stuff flows across the branches of the U and to be able to do that with code simply and quickly was something that the functional APIs were designed for, so by covering them in the first two courses of this specialization, we're now able to start digging into things like a unet and I find that really exciting where you have perhaps I think the most inspirational part for me though, is really when we start looking at the interpretability of the model using things like the class activation maps and salience and there's one really fun example that we go through is if you recall, in the earlier convolutional neural networks courses, we had an exercise to determine cats versus dogs to build a convolutional neural network to be able to classify those. We explored image augmentation because at the time when we were building that we're thinking, well, a lot of the times we can only classify a cat because we're maybe looking at the ears, the very distinct ears of a cat, those triangular pointy things, but maybe some cats are sideways and we may not spot it as a cat if that's the case, but when we start applying the salience and class activation maps that we learned about in this course, we ended up realizing that the one thing that really made the distinguishing factor between a cat and dog were the eyes and when you start seeing like the heatmaps lighting up the eyes of the cat to say, we know this is the cat because of these heatmaps, I just find it's like, it's really cool to see that this is what the neural network was paying attention to, and from that it was able to do that determination and now how can we use that then to see how computers are visualizing things so that we can maybe understand our models better. We know both from how people recognize other people and from face-recognition well that in recognizing humans the eye is actually one of the most distinctive cues. We don't talk about it because we don't have a vocabulary really to describe people's eyes and I think is really cool that it cannot allows you to discover that even for recognizing cats versus dogs. The eyes turns out to be where the new unethical pays attention. Absolutely. Also like on the theme of making decisions as an engineer is one of the things that I really enjoy in this course is some of the APIs that we look into to help us understand the output of the model. For example, for object detection models, it's pretty common that the model gives you a lot of data, so it might be spotting a number of instances of an object in an image, and then that instance has a probability of a particular label and then there were bounding boxes for that and then there were multiple ones of these. You have to write a lot of code to be able to pass through all of that. One of the things that we do, look at it in the course or the object detection APIs, which as well as doing object detection, they also hold your hand a little bit through that so that instead of writing a hundreds of lines of code to determine which bounding boxes do you really want to be interested in? That it really helps you do that by managing that for you, loading this stuff into dictionaries and making it easier for you to do those plots. I think what I've say when learning one of the most interesting things was how do you structure a neural networks outputs. What do you want a neural network to output so that it's not just classifies an object that gives you all the information you might need to figure out the bounding box around it, so the Tensor Flow APIs mix all that easier to just code up and deal with. Yeah, so, and will have an exercise of the course where we build a zombie detector using these so that we have images of a zombie walking along the road and we're going to use those APIs to draw a bounding box around the zombie, like determine what's a zombie and what's a door or other things in the background and you never know it might be handy one day. It is ever a zombie apocalypse. I think a lot of learners really glad they learn about these tools and can use them to help protect us all from this terrifying zombie hoard. This would be an exciting set of materials to be covering. Here we go into advanced computer vision.
