Human vision has evolved to recognize human figures in all different situations, whether it is the real world or art through representations.
The situation becomes a little more complicated when more abstract art forms such as cubism or surrealism are involved, in which the shapes are distorted, often making their recognition difficult. If humans have proven that they have no difficulty recognizing a person in a Picasso painting, for computers things do not seem to be so simple.
This question, namely, to what extent a computer program can recognize human figures in paintings belonging to the Cubist movement, was attempted to be answered by some scientists from the University of Berkeley in California, USA, in their study "Recognizing People in Cubism".
Cubism was one of the most influential movements in 20th-century painting, with pioneers such as Georges Braque and Pablo Picasso. In the images of the corresponding paintings, the figures are presented distorted, as if the perspectives that observers would have from different angles were united. As a result, a Cubist painting contains many fragments of the painter's perception of the same object.
The researchers used a list of 218 Cubist paintings with titles that indicated they contained human figures. They then asked 18 participants to rate the degree of abstraction in each image on a scale of 1 to 5 and to draw a rectangle around each person.
They then turned to their computers, trying to replicate the results. To do this, they used algorithms based on neural networks, a technique that in recent years has revolutionized the field of facial recognition as it uses existing data to train and improve its performance. Initially training their system with photos of people, they then fed it with the database of 218 paintings that they had presented to the research participants, with the same requirement to place an outline around each person.
The results provide a very useful insight into how computer vision algorithms work, while also revealing insights into human perception. Unsurprisingly, humans do better at estimating cubism, scoring twice as high on the relevant assessment as computers. In fact, as the abstraction in the table increases, the results for the algorithm become increasingly disastrous, unlike humans, who do not seem to be affected as much. This observation contradicts a widespread belief among neuroscientists that humans perceive objects by analyzing their individual parts.
It seems, however, that if computer vision can compete with and even surpass human vision in normal conditions, the abstract conditions that art subjects them to show that human perception is much more flexible and reliable. However, the results of this specific research will be used to optimize the vision of computers, which one day not too far from today, may be able to better appreciate a modern art painting.
Source: naftemporiki.gr
