2014 CVPR Histogram Dirichlet-based Feature Transf
no vote
Application background With the Dirichlet distribution of the new vector Fisher implementation, compared to the original vector Fisher training parameters to reduce, the origin of the three into one, improve the efficiency of training, but also improve the accuracy of training. Key Technology Development of image classifications such, as by recentLocal descriptors. In this paper we, propose a method SIFTEfficiently transform those histogram features for improving toClassification performance. The the (L1-normalized)Feature is regarded as a probability mass function histogram,Is modeled by Dirichlet distribution. Based on whichProbabilistic modeling we, induce the Dirichlet Fisher theFor transforming the histogram feature vector. The kernelWorks on the individual histogram feature to enhance methodDiscriminative power at a low computational theOn the other hand in, the bag-of-feature cost. (BoF) framework,Dirichlet mixture model can be extended to