A vocabulary-tree implementation for mobile product recognition

In this work we present a method for mobile image search and retrieval. Local image attributes are modelled by using SIFT features and codewords are obtained by using the Bag-of-Features (BoF) approach. Extracted features and generated codewords are indexed within a vocabulary tree (VT). Matching is performed by using a weighted Euclidean similarity criterion which is defined according to the VT structure and is robust to distortions such as rotation, changes in illumination, cropping and scaling. An XML based structure is proposed in order to represent the VTs in the database. Experimental results are obtained by using “Stanford Mobile Visual Search Data Set” and reported alongside with the results obtained via exhaustive search. It is shown that the system provides around 90% recognition rate with a low computational complexity thus is suitable for mobile applications.