Abstract:When codeword frequency meets geographical location in landmark search applications, is it still discriminative for the search procedure? In this paper, we give a systematic investigation about how geographical location affects the effectiveness of codeword frequency. We explain why the standard IDF in the BoW models is less effective in location related search applications [11][12]. Consequently, we propose a "location discriminative codeword frequency" strategy to introduce the location context into the codeword discriminability measurement. This new codeword frequency is calculated in each geographical region, for which a spectral clustering scheme is proposed to partition the geographical map of each city into distinct regions. Extensive comparisons over the standard codeword frequency in state-of-the-art landmark search systems [1][1] demonstrates our approach's effectiveness.

When Codeword Frequency Meets Geographical Location.