How to see like a computer
Aug 17, 2015
Our brains work daily to assemble shapes and discern object using location-specific neurons in V1, primary visual cortex. V1 is not a code name for a robot, actually it’ss the best-studied visual area of human brains, which is specialized for processing information about static and moving objects, and pattern recognition. Therefore, it’s not strange that our computer vision techniques started on the foundation of human visual experience.
David Hubel and Torsten Wiesel, Nobel-prize winners, proposed an organization model which includes two tuning properties – ocular dominance and orientation. Nowadays, research into biological vision enters the areas of color, spatial frequency and other features which are still not accompanied by this model. However, we don’t rely only on human experience: different computational models help us in various realizations of artificial intelligence. AI often deals with robotic navigations through different environments, which is usually provided by various computer vision systems. These systems rely on various image sensors, and accompany lots of data from quantum physics and neurobiology.
So, can a computer outperform humans? Of course it can, computers can discern various textures and shades that humans cannot, and computational visual methods are great from detecting forgeries (for example by analyzing brushstrokes of famous artists) and small dissimilarities in images, that humans fail to see. However, a question jumps out of the crowd: can a human outperform a computer?
People are good at picking out statistical irregularities and global overall detection, and they are good at solving NP-complete problems. The latter are a class of decision problems (that seek a yes/no answer) which cannot be efficiently solved by a computer, but we can approximate various solutions. For example, a computer would try out every possible combination if we were told that there are 100 pigeon holes and 101 pigeons, but a human would see a general principle behind it. That’s the same technique that astrophysics uses nowadays, since people are better at noticing some visual patterns, rather than their robotic counterparts. For example, Planet Hunters uses human volunteers (robots can apply too!) to classify brightness curves of planets that did their transit in front of a star, which led to a small, but visible decrease in its brightness.
But if we’re so great, why do we need computer vision? Well, for starters, Mars is currently the only planet solely inhabited by robots! Computer vision mechanisms are extremely useful for rovers and robotic landers, to analyze different data and move around unknown surfaces with ease. That would not be possible if our robot was to bump into rocks every now and then, so computer vision and tracking algorithms are used to create optimal paths.
Okay, that’s great, but I need to know how can I use tracking and computer vision techniques on Earth, rather than on Mars? Computer vision as a field of recognition, which includes object recognition, detection and identification, which is used for security identifications, but in modern research in biology and medicine, especially for early detection of various abnormalities and disease, that people would not be able to do. Various feature extraction algorithms can be of use to marketing retailers and researches in various scientific disciplines like psychology, to detect emotions and other data from sets of images or videos. More on, assistive technologies nowadays incorporate various tracking algorithms to help people with disabilities to use computers and mobile devices easily. Even when you tag your friends on Facebook or Instagram, you’re dealing with computer vision algorithms, that include facial recognition and object recognition techniques.
A new industrial trend is the concept of augmented reality, which enables you to experience the world around you using an indirect and improved view. Applications of these techniques enhance one’s perception of reality, and can be used to try out different products, to embody different virtual characters, and to interact with the world which becomes digitally manipulable.
Recently, computer vision techniques have advanced into computational aesthetics as well, returning abstract concepts of beauty or symmetry from images, so our advantage over computer may slowly be decreasing. Computers are able to predict moods or emotions, which is a quality lots of us humans fail to do in everyday life. The field of computer tracking and vision research is still taking off, but when it finally realizes its potential, the results will be stunning.
- Medioni, Gérard and Sing Bing Kang (2004): Emerging Topics in Computer Vision. New Jersey: Prentice Hall.
- Orbán LL and Chartier S (2015): “Unsupervised Neural Network Quantifies the Cost of Visual Information Processing”, in: PLoS ONE 10(7): e0132218.
- Redi et al. (2015): “Like Partying? Your Face Says It All. Predicting the Ambiance of Places with Profile Pictures“, at: arxiv.org