2012-09-23

Keywords:

• large scale image datasets
• rootSIFT
• image augmentation
• query expansion
• paris buildings

Q1: What's RootSIFT and how does it improve over L2?

/RootSIFT/ is a modified SIFT descriptor where the elements are square roots of L1 normalized SIFT descriptors. Comparing RootSIFT descriptors with Euclidean (L2) is equivalent to using Hellinger kernel to compare SIFT. Hellinger kernel is $d_E(\sqrt{x}, \sqrt{y})^2 = 2- 2 H(x, y)$.

It improves over Oxford 105k baseline system from 0.515 to 0.583 in tf-idf ranking.