Training Deep Models and Deriving Fisher Kernels: A Step Wise Approach

Abstract

Deep networks and Fisher kernels are two competitive approaches showing strides of progress and improvement for computer vision tasks in specific the large scale object categorisation problem. One of the recent developments in this regard has been the use of a hybrid approach that encodes higher order statistics of deep models for Fisher vector encodings. In this chapter we shall discuss how to train a deep model for extracting Fisher kernel. The tips discussed here are validated by industrial practices and research community through mathematical proofs by LeCun et al. (Neural networks: tricks of the trade. Springer, pp 9–50 (1998), [1]), Bengio (Neural networks: tricks of the trade. Springer, pp 437–478 (2012), [2]) and case studies.