Active Object Recognition with a Space-Variant Retina

Figure 13

Our approach’s mean per-class accuracy on AR and Caltech-256 as a function of the number of test fixations. For AR, 8 training instances per category were used, and for Caltech-256, 50 training instances per category were used. The results are averaged over five cross-validation runs. The black dots indicate the number of fixations needed to achieve 95% of the maximum accuracy (17 for AR and 152 for Caltech-256).