Table 1.

Confusion matrices for the CalTech-4 dataset with one multicategory classifier

	Classifier $\to$ (SUVM + SVM)/[Fergus et al. (15)]
Query $\downarrow$	(F)ace	(M)otorbike	(A)irplane	(C)ar
F	0.982/0.862	0.000/0.073	0.018/0.028	0.000/0.014
M	0.000/0.000	0.990/0.977	0.010/0.013	0.000/0.000
A	0.005//0.003	0.013/0.042	0.967/0.888	0.015/0.060
C	0.000/0.008	0.000/0.092	0.020/0.197	0.980/0.670

The table entry (i, j) is the percentage of query images belonging to category (i) that are classified as belonging to category (j). Each table entry is separated with the delimeter “/”: the number to the left is the performance of the (SUVM + SVM) approach and the entry to the right is the performance for Fergus et al. (15). For SUVMs, a visual dictionary is learned from all of the images (i.e., a shared visual dictionary is created), but each model is learned only from its category-specific images. A single 4-class SVM classifier is built by combining the outputs of all of the four models as was done in ref. 15.