Gaussian Normalization: Handling Burstiness in Visual Data
Refereed Conference Meeting Proceeding
This paper addresses histogram burstiness, defined as the tendency of histograms to feature peaks out of pro- portion 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 intranormalized Fisher vectors.
IEEE International Conference on Advanced Video and Signal-based Surveillance (AVSS)
Digital Object Identifer (DOI):
Taiwan, Province of China
Dublin City University (DCU)
Open access repository: