Figure 6.

Classes of animal features were decomposed in a multidimensional space using principal component analysis. For each class, the distance with other classes correspond to the proportion of common images between the two classes. All classes have thus coordinates in a multidimensional space and PCA extracts the two first principal dimensions. This representation makes it possible to estimate distance between classes (i.e. the rate of co-occurrence of image characteristics). Complementary classes of characteristics (“partially visible”/“Totally visible” for example) are symmetrical compared to the origin in this space. Very clearly four groups, manually separated in dotted lines, are profiled. This representation only depends on image statistics.