Classifier performance depends on picture similarity

In general, stimulus discriminability in IT (and thus classifier performance) depends on similarity (e.g. it is harder to separate a face from another face than a face from a car). To further quantify this observation, we evaluated the classification performance in identification within each of the 8 individual groups. This showed that, for example, individual pictures within the toys or foodstuff groups were easier to discriminate than pictures within the vehicles group or the monkey faces group. Not surprisingly, the set of white box-like shapes showed the worst within-group classification performance. The computational difficulty of any visual classification task depends on image similarity but also on the recognition architecture (for instance on the dictionary of features that may be used by it).

Read-out performance as a function of the pixel similarity among pictures within a given group.