Query Types and Visual Concept-Based Post-retrieval Clustering

In the photo retrieval task of ImageCLEF 2008, we examined the influences of image representations and clustering methods for enhancing instance recall when they are used in post-retrieval clustering. Two types of visual concepts and hierarchical and partitioning clustering methods were compared. We used the title fields in the search topics, and either only the title fields or both the title and description fields of the annotations were in English. The experimental results showed that one type of visual concept representation dominated the other except for in one condition. Also, it was found that hierarchical clustering can enhance instance recall while preserving the precision when the threshold parameters are appropriately set. In contrast, partitioning clustering degraded the results. We also categorized the queries based on the clustering criteria, geographical and non-geographical, and found that the geographical queries are relatively easy and post-retrieval clustering works better for them.