Gaussian Normalization: Handling Burstiness in Visual Data

Abstract

This paper addresses histogram burstiness, defined as the tendency of histograms to feature peaks out of proportion with their general distribution. After highlighting the impact of this growing issue on computer vision problems and the need to preserve the distribution information, we introduce a new normalization based on a Gaussian fit with a pre-defined variance for each datum that suppresses burst without adversely affecting the distribution. Experimental results on four public datasets show that our normalization scheme provides a staggering performance boost compared to other normalizations, even allowing Gaussian-normalized Bag-of-Words to perform similarly to intra-normalized Fisher vectors.