Search

We describe a computational model of humans' ability toprovide a detailed interpretation of a scene’s components.Humans can identify in an image meaningful componentsalmost everywhere, and identifying these components is anessential part of the visual process, and of understanding thesurrounding scene and its potential meaning to the viewer.Detailed interpretation is beyond the scope of currentmodels of visual recognition. Our model suggests that this isa fundamental limitation, related to the fact that existingmodels rely on feed-forward but limited top-downprocessing. In our model, a first recognition stage leads tothe initial activation of class candidates, which isincomplete and with limited accuracy. This stage thentriggers the application of class-specific interpretation andvalidation processes, which recover richer and moreaccurate interpretation of the visible scene. We discussimplications of the model for visual interpretation byhumans and by computer vision models