Introduction. We consider the problem of binary hashing, where given a high-dimensional vector x ∈ R D , we want to map it to an L-bit vector z = h(x) ∈ {0, 1} L using a hash function h, while preserving the neighbors of x in the binary space. Binary hashing has emerged in recent years as an effective technique for fast search on image (and other) databases. While the search in the original space would cost O(N D) in both time and space, using floating point operations, the search in the binary

more »

... space costs O(N L) where L ≪ D and the constant factor is much smaller. This is because the hardware can compute binary operations very efficiently and the entire dataset (N L bits) can fit in the main memory of a workstation.