Active Object Recognition with a Space-Variant Retina
Figure 12
Mean per-class accuracy for our approach on Caltech-256 as a function of the number of training instances compared to the methods of [10, 26, 33, 36, 37]. Chance performance is . Kanan [26] used a gnostic field with color SIFT features, and our space-variant ICA filters achieve almost the same accuracy (slightly more for one training instance), despite being a self-taught approach. Bergamo and Torresani [37] combined five kinds of features (color GIST, oriented HOG, unoriented HOG, SSIM, and SIFT) into a metadescriptor using spatial-pyramid histograms. Gehler and Nowozin [36] used five types of engineered features (PHOG, SIFT, LBP, V1+ Gabors, and region covariance) and used multiple kernel learning to combine 39 different kernels. Kanan and Cottrell [10] used a nonfoveated model of active vision (see discussion). Griffin et al. [33] provides baseline results.
