Week 4: Computer Vision

This time I decided to pick a topic I'm familiar with (a few years of professional work in the industry), called Computer Vision, although some of you have some significant knowledge of it. Most people have very divergent opinions. Computer Vision is a very young part of IT field and computer knowledge. In Poland there are almost no specialized CV departments. In USA there are separate faculties founded with vision of researching... computer vision. Nevertheless vision revolution is behind the corner.

The growth of this particular field was possible thanks to technological progress in the imaging market (high resolution cameras, storage of thousands of pictures and movies) and embedded systems (high computing capabilities in our phones, cars, cameras, watches, even fridges....). In fact, that progress is not really connected with hardware. Hardware is the Opto-electronics and Electronics market. Computer Vision uses many medical researches about cognition of human. After extraction whichever data it needs from an image, CV algorithms try to outperform a human in recognizing, analyzing and categorizing images. It's really a simple idea at first sight. If there is a chair in the image, it's just natural to write a program that says so. In fact …. recognizing a chair is a big problem for engineers. A human can not only perceive on hundreds levels of recognition layers, but can also create non-existent data. Our biologic vision reconstructs most of the data we need to see a chair. I will try to bring the topic closer to you with some real life applications that are used now or will be used in a few years.

"What is It?" nr 1.

"What is It?" nr 2.

"What is It?" nr 3.

First big wave of modern computer vision came with algorithms of face detection. Nowadays cameras have it, facebook has it, every intelligent CCTV camera has it. Face detection is the first so successful technology to come into houses of common people. Of course, even before facebook there were created military-grade systems for tracking and intrusion detection but even the best systems had to be controlled by humans. The false positive rate was too high and cost of such systems was huge because of low computing capabilities of computers. The industry had it, though it was (and in many fields still is) limited to very simple operations as blob detection and color segregation. Unfortunately industry needs success rate or efficiency of visual cognition systems on level higher than 99.9% and it cannot be achieved in most cases with current state of the art algorithms. So the biggest market for computer vision became mass market, where each new visual function is well received even if it fails in 20% cases. There are technologies that are nowadays seen as a computer vision but are really rather optical systems or sensors, as laser barcode scanners.

Why face detection? Because face is so easy to detect. It's so characteristic and different in nature and synthetic graphics (made by a human), that its detector has a high detection rate even in the dumbest implementations seen in first “intelligent” cameras. In comparison, there is no known descriptor for a mug ... In fact people have problems with it as well. They recognize environment, as background: table, kitchen, spoon etc. and assume that something of such a simple shape has to be a mug. Without background it is a very hard task even for a human.

Detection and recognition are two different tasks. If we can detect a face...we can detect only a part of graphics that looks like a face. We don't know that it's “Brad” or “Eva”. That technology is just starting to evolve and you can see it on facebook, though they are striving to upgrade the system for high recognition rate. Its roots are in biometrics, which you can probably know from science fiction films and fingerprints and iris scanners.

So face detection is really a simple task. If you want to analyze images of buildings, crowds, cars, different actions and objects, moving bodies – it's a whole new world to research. So what can we really observe nowadays in the state of the art computer vision industry?

These are the color-normalized images of more than 1000 images of the same object in different scenes. Even after such generalisation, the scene has critical information needed for recognition task. The information in sole object is not sufficient in most cases.

Let's start with surveillance. You probably know that London has hundreds of thousands CCTV cameras already installed. Each London citizen is recorded more than a 300 times a day. At first it was used mainly to track down suspects “manually”. What can do computer vision do nowadays? These are the data that can be extracted in real-time from a camera for business intelligence purposes;  for example, people crossing a scan line, baggage, car  tracking,  sitting, walking, running.  And that's only CCTV. Do you know what data some companies take from cameras in big malls?  Places of biggest attraction, speed of walking between specific shelves. Some cameras can say where you are moving your hand while taking a product from a shelf. In most cases such systems have low efficiency, at 50-70%.

On the other hand , U.S. require new cars to have rearview cameras by 2018. Nvidia created some time ago a new embedded system with huge computational powers suited for cars (Tegra TK1) for that reason. The car accidents will be recorded with exact information about car numbers, gps data and environmental specification. If there are cameras on the side and front view, the systems for automatic parking will be commonly used. Best vision systems (Google has been a pioneer for a long time) can drive a car better in a city with many people than a real driver -

http://www.technologyreview.com/news/520746/data-shows-googles-robot-cars-are-smoother-safer-drivers-than-you-or-i/  .

Can you see a car and a pedestrian? This pedestrian is a car graphics copied, rotated and moved. Do we want to build an algorithm that "thinks", that there is a pedestrian or a rotated car? 

Computer vision is 3D imaging as well. These are 3D scanners with structured light, Time of Flight (ToF technology), laser and infrared stuff that most of common people don't know. We can do 3D models with a single lens camera. So now...imagine that we can make a sparse 3D model of each town. Looks like Google Street View? Nope. Google wants to create real 3D towns. And now it can be constructed from a mere.... Flickr photos.

https://www.youtube.com/watch?v=ofHFOr2nRxU

OCR technology is almost everywhere. We can build document scanners (as a passport scanner) with a simple internet camera and some programming skills. We can analyze whole graphics and tell what elements make it attractive to people.

Google indexes all graphics on the internet with many levels of graphical descriptors. If you use some licensed graphics on your website – Google knows. Even if it was resized, cut in half, compressed and color space was altered. In USA if you earn a lot of money using such graphics, you will meet the police department quicker than you bought your server.

Did you know that with a single smartphone camera you can measure your own heart rate? That's called micro moves. Just watch it on YouTube. That's a better introduction than reading about it...

https://www.youtube.com/watch?v=3rWycBEHn3s