Packed dense interest points for scene image retrieval

In this paper we propose a novel method to retrieval scene images, where packed dense interest points are extracted as visual signature and asymmetric piece-to-image matching scheme is adopted for similarity matching. Compared to general images, scene image contains large area of low contrast regions, such as sky, calm water, or flat road surface, which make it difficult for popular method to extract appropriate representation with both good coverage of the entire image and insures relatively high repeatability. To obtain appropriate representation of a scene image as its visual signature, our method extracts packed dense interest points, i.e. using dense interest points as point features then packing these features in groups according to their spatial relations. After that a piece-to-image matching scheme based on packed Dense Interest Points is applied to fulfill retrieval work. Our retrieval method enhances interpretation ability of signature for scene image and performs matching taking geometric constrains into account. We evaluate our approach on MIT 8 scene categories, and demonstrate clear improvements in retrieval over conventional purely appearance-based baselines.