Stopwords Detection in Bag-of-Visual-Words: The Case of Retrieving Maya Hieroglyphs

We present a method for automatic detection of stopwords in visual vocabularies that is based upon the entropy of each visual word. We propose a specific formulation to compute the entropy as the core of this method, in which the probability density function of the visual words is marginalized over all visual classes, such that words with higher entropy can be considered to be irrelevant words, i.e., stopwords. We evaluate our method on a dataset of syllabic Maya hieroglyphs, which is of great interest for archaeologists, and that requires efficient techniques for indexing and retrieval. Our results show that our method produces shorter bag representations without hurting retrieval performance, and even improving it in some cases, which does not happen when using previous methods. Furthermore, our assumptions for the proposed computation of the entropy can be generalized to bag representations of different nature.