How Computer Vision Algorithms Cope With Detecting Human Fig
Source: KentuckyFC
The human visual system has evolved to recognize people in almost any pose under a vast range of lighting conditions. But abstract art pushes this ability to its limits by distorting the human form. In particular, Cubism seeks to represent three-dimensional objects on a two-dimensional plane by juxtaposing snapshots from different angles. The result is that a Cubist picture contains many 'fragments of perception' of the same object. That's why it is often hard for people to recognize the human figures that these pictures contain. Now a group of computer scientists have tested how computer vision algorithms fare at the task of spotting human figures in Cubist art. They compared a variety of different algorithms against humans in trying to spot human figures in 218 Cubist paintings by Picasso. Humans easily outperform all the algorithms at this task. But some algorithms were much better than others. The most successful were based on so-called "deformable parts models" that recognize human figures by looking for body parts rather than the entire form. Interestingly, the team says this backs up various studies by neuroscientists suggesting that the human brain works in a similar way.
| }
|