Computer Vision News - September 2016

Rijnders found that on the level of neurons, our brain develops models like the chair and table in the first years of our lives: our consciousness doesn’t process all of the single photons. Instead, our consciousness stores these models for faster recognition and processes models that have been recognized by our visual cortex. Then we make a conscious model out of this. That’s the inspiration behind his idea. He soon realized that computer vision would need much more refined algorithms . As an engineer, his philosophy is that the underlying data has to be as good as possible. Not only that, but you should give the engineers a model that they can understand. That is how Cogisen is becoming very relevant in areas like machine learning and deep learning , which are becoming very effective but are still limited by the underlying data. Here comes the idea of gaze tracking, one of Cogisen’s applications in computer vision: you can’t just provide a deep learning model of many faces with their eyes moving. First of all, you must have a relevant model of the eyes moving. Secondly, you have to be able to do things sparsely and in a non- linear way. You have to consider the infinite number of light conditions, the faces, and the points of view to the camera. Then you consider the fact that the pupil and iris movements are subpixel. The movements of the iris don’t even register in pixels if one is standing one meter from the camera. Think about the area on the inside of the eye, the outside of the eye and the irises. You can’t measure these movements by counting the pixels unless under perfect conditions. Instead of counting pixels, Cogisen makes the underlying data much better and creates a model from it . With this work, difficult nonlinear problems in industry can find a solution: for example, the detection of distant objects for autonomous vehicles with very few pixels. This could also include gaze tracking interfaces for people standing meters away from the video as well as adaptive video compression where it is difficult to get a good model of visual saliency. Currently, the most advanced solutions in image processing need heavy GPU use. If you look at what they are doing, such as the cloud image processing or image processing for autonomous vehicles, it is extremely advanced and computationally intensive. This cannot be the solution for the real needs of the industry which is Internet of Things . We will have billions of devices for which such a huge amount of GPU usage is out of question. Computer Vision News Application 9 “ The movements of the iris don’t even register in pixels if one is standing one meter from the camera ” Application