Hyperspherical embedding iterative quantization hashing

The iterative quantization hashing learns similarity preserving binary codes by rotating the data points in a space projected by the principal components of data to minimize the quantization error between hash function outputs and corresponding binary codes. However, in some cases, a rotation may not be enough to preserve similarity by Hamming distance. In this paper, we propose a hashing method that embeds the data points on to the surface of a hypersphere according to the data similarity before performing the iterative quantization. Experimental results show that the hash codes learned by the proposed method approximate the neighborhood relationship with higher precision.