 |
Center for Nonlinear Studies |
In the computer vision field, performance is judged by the accuracy of classification on well-known image sets. However, these image sets might not be a good proxy for vision, due to several factors, like outside information. In this talk, I will discuss the importance of picking a correct image set for testing computer vision and I will introduce a novel image set for testing a V_1 system. I will then show that even though humans can easily classify it, several current feedfoward systems cannot do better than chance. I will show, however, that if we allow lateral connections in our models, we can find a connectivity kernel that will increase the performance up to human psychophysics levels.
Host: Garrett Kenyon