On Building a Universal and Compact Visual Vocabulary

Figure 1

Recognition rates of different similarity thresholds. With all three datasets, the similarity thresholds 0.75 and 0.8 produce the best or near-best performance.