The accuracy and value of machine-generated image tags: design and user evaluation of an end-to-end image tagging system

Automated image tagging is a problem of great interest, due to the proliferation of photo sharing services. Researchers have achieved considerable advances in understanding motivations and usage of tags, recognizing relevant tags from image content, and leveraging community input to recommend more tags. In this work we address several important issues in building an end-to-end image tagging application, including tagging vocabulary design, taxonomy-based tag refinement, classifier score calibration for effective tag ranking, and selection of valuable tags, rather than just accurate ones. We surveyed users to quantify tag utility and error tolerance, and use this data in both calibrating scores from automatic classifiers and in taxonomy based tag expansion. We also compute the relative importance among tags based on user input and statistics from Flickr. We present an end-to-end system evaluated on thousands of user-contributed photos using 60 popular tags. We can issue four tags per image with over 80% accuracy, up from 50% baseline performance, and we confirm through a comparative user study that value-ranked tags are preferable to accuracy-ranked tags.