This talk will focus on the problem of learning similarity-preserving binary
codes for indexing and search of large-scale image collections. The idea is to
map high-dimensional feature vectors to compact bit strings such that vectors
representing semantically or perceptually similar images in the original
feature
space map to strings with a low Hamming distance. This talk will describe
Iterative Quantization, a simple and efficient binary coding scheme that works
by finding a rotation of zero-centered data to minimize the quantization error
of mapping the data to the vertices of a binary hypercube. We will also
describe
an application of similarity-preserving binary codes to clustering and
reconstruction of landmark and city photo collections.