On Building a Universal and Compact Visual Vocabulary

Table 1

Classification rates of different vocabulary sizes on three datasets. The sizes corresponding to similarity thresholds 0.75 and 0.8 produce the best or near-best results in all cases.

	caltech-101	Event-8	Scene-15
100	36.6 ± 1.0	70.7 ± 2.5	65.0 ± 0.4
1000	48.5 ± 1.1	79.3 ± 2.1	73.1 ± 0.7
10000	48.3 ± 1.1	80.0 ± 2.2	73.9 ± 0.5
50000	47.1 ± 0.9	78.5 ± 2.1	72.1 ± 0.5
100000	46.5 ± 1.1	78.1 ± 2.0	71.3 ± 0.7
th𝑐=0.75	49.5 ± 0.9	79.5 ± 2.4	73.9 ± 0.6
th𝑐=0.8	49.0 ± 1.0	80.4 ± 2.5	73.8 ± 0.5