[SOUND] Hello, first I want to make a brief introduction to the field of computer vision. What is computer vision? To answer this question, we need to understand first what human vision is. There is a saying, to see is to know what is where by looking. The goal of computer vision is to solve the same problem on a computer. Thus, computer vision is regarded as a part of artificial intelligence. In computer vision we have our own version of Turing test. To pass it, the computer vision system should answer any question regarding an image, which can be answered by a human. What are the main questions about an image? The first question comes from the informal definition of vision. This, what and where? What objects are seen in the image and where are there are located? This is called object detection. Some objects are more important for us than other. Because we're humans, the most important object for us are other humans and their faces. Thus, it is understandable why face detection and person detection are one of the most rigorously researched topics in computer vision. And still they are not solved completely, with new algorithms appearing every week. In this particular image we can see other important objects like text, or less important, buildings and flags. After objects are detected, we move on to know their properties. Is it the picture of a city, or it is something else? What color is the sky, how strong is the wind? What can we say about this person? Currently these problems are usually solved by image classification. Possible answers are enumerated and an algorithm predict the correct answer. For example, we can divide all images into images of cities and all the rest, and then it classify as city or not city. To identify the particular object and person, we usually store the images and compare the current image with the stored, which is called image retrieval. We can also obtain metric information from images, like distances between selected points in scene. We can also estimate the shape of the object, position of camera relative to the object, distance from the camera to the object. Currently available algorithm of metric vision allows automated creation of fully textured 3D models from set of photographs downloaded automatically from the Internet. The quality of these models are not of production quality yet, and you can't directly put them into computer game, but we will get there soon, I believe. In our human brains we combine our observations with prior knowledge regarding the volt. It is estimated that a quarter of human brain is devoted to the vision. A lot of computer vision algorithm have derived ideas from our knowledge of human vision system. Probably the most influential work computer vision is the model of human vision system, which was proposed by David Marr in 70s. It consists of three steps. First, low-level image features, like directed edge segments, are extracted from image to form so-called primal sketch. From this primal sketch a two 2.5D sketch is obtained using binocular stereo. Finally, high-level techniques using prior knowledge is used to get 3D model representation of objects in the scene. This seminal work has received so much appraised that the most prestigious award in computer vision is named after the Marr, the Marr Prize. This is awarded each couple of years during ICCV, International Conference on Computer Vision. The progress in computer vision is entangled with progress in machine learning. As an example, I want to mention the seminal work, the Viola-Jones face detector, which was proposed in 2001. It was the first detector which can elaborately detect human faces in real time on a personal computer. This was a demonstration of power of machine learning and data collection. Until recently, if you saw a face detector in a consumer product, it was definitely a Viola-Jones detector or one of its derivatives. Neural networks has a long history in computer vision. But since 2012, when AlexNet was introduced, we are witnessing a record improvement of computer vision algorithms. Deep learning has been successfully applied to most of the computer vision problems. In this course we will focus on deep learning methods in computer vision. But we will also introduce a lot of general computer vision concepts, explain both basic and seminal non-deep learning methods, talk about contemporary data sets. All to give you perspective and understanding of key ideas, problems, and achievements of computer vision. [SOUND]
