Soft assignment vs hard assignment coding for bag of visual words

In image classification and retrieval, the semantic gap is the major challenge. It characterizes the difference between human perception of a concept and how it can be represented using machine level language. Bag of visual words is a well-known efficient method for image representation, however it showed some limitations. The loss of information during the vector quantization process is one of these limitations. Many approaches were proposed in order to deal with this, such as the soft-assignment technique and sparse or local coding schemes. This paper aims to compare and evaluate the extent that has the soft-assignment approach especially the recently proposed locality-constrained linear coding over the traditional hard assignment methods.