Nonparametric Estimation of Fisher Vectors to Aggregate Image Descriptors

We investigate how to represent a natural image in order to be able to recognize the visual concepts within it. The core of the proposed method consists in a new approach to aggregate local features, based on a non-parametric estimation of the Fisher vector, that result from the derivation of the gradient of the loglikelihood. For this, we need to use low level local descriptors that are learned with independent component analysis and thus provide a statistically independent description of the images. The resulting signature has a very intuitive interpretation and we propose an efficient implementation as well. We show on publicly available datasets that the proposed image signature performs very well.