Fig. 9

Rotation invariance when training on two orientations. Plot of mean classification accuracies (in %) when training on the two orientations θ∈{0°,40° } and testing on θ∈{80°,120°,…320° } for each descriptor applied on textures rotated by hardware (HW) or nearest neighbour (NN), linear (LN), cubic (CU), B-spline (SP) and Lanczos 3 (LZ) interpolation. The horizontal red dashed lines are the mean accuracies when training and testing on θ=0°. The grey dots in the background correspond to the mean classification accuracies achieved when the size of the training data is halved. The horizontal line segments mark the median in each distribution