Caption vs. Query Translation for Cross-Language Image Retrieval

For many cross-language retrieval tasks, the predominant approach is to translate the query into the language of the document collection (target language). This often gives results as good as, if not better, than translating the document collection into the query language (source language). In this paper, we evaluate query versus document translation for the ImageCLEF 2004 bilingual ad hoc retrieval task. Image retrieval is achieved through matching textual queries to associated image captions for the following languages: French, German, Spanish and Italian using commercially and publicly available resources. On average, we find query translation to outperform document translation (77% of English MAP compared to 65% respectively) but this varies widely across language and query. Combining document and query translation we achieve an average MAP of 85% of English.