Boosting landmark retrieval baseline with burstiness detection

Abstract

In image retrieval, the bag‐of‐visual‐words model‐based approaches combined with the spatial verification (SP) post‐processing step have achieved considerable progress. However, in practice, especially for retrieving landmark images, the authors have observed that this baseline suffers from the problem of burst matches. This issue is caused by repetitive visual patterns that appear frequently among images. Local features derived from these burst patterns can redundantly match others, resulting in many invalid matches that vote over‐estimated similarity scores for irrelevant images. Essentially, this problem can be mainly attributed to two reasons, (i) the non‐exclusive matching leads to one‐to‐many matches, (ii) the SP fails to filter burst matches that are closely located. To tackle this problem, a burstiness detection approach using geometric and visual word information of local features is proposed. Firstly, a geometric filtering strategy is employed to remove matches that are not consistent with global scale variation. Then, the one‐to‐one matching strategy is applied to detect and eliminate one‐to‐many matches. Finally, a down‐weighting burstiness strategy is adopted to penalise the voting weight of burst matches. Experimental results on three public datasets demonstrate that the proposed approach can achieve a comparable or even better accuracy over other popular approaches.

Keywords

• boosting landmark retrieval baseline
• bag-of-visual-words model-based approaches
• spatial verification post-processing step
• SP
• burst matches
• burst patterns