Effects of Visual Concept-based Post-retrieval Clustering in ImageCLEFphoto 2008

We examined the effectiveness of post-retrieval clustering that was based on the visual similarities among images to enhance the instance recall in the photo retrieval task of ImageCLEF 2008. The visual similarities are defined by the example visual concepts that were provided for the automatic photo indexing task. We tested two types of visual concepts and two kinds of clustering methods, hierarchical and modified k-means clustering. In all the runs, we used only the title fields in the search topics; we used either only the title fields or both the title and description fields of the annotations in English. The experimental results showed that hierarchical clustering can enhance instance recall while preserving the precision when certain parameters are appropriately set.