User evaluation of automatically generated keywords and toponyms for geo-referenced images by Frank O. Ostermann, Martin Tomko, Ross Purves. (Ostermann, F. O., Tomko, M. and Purves, R. (2013), User evaluation of automatically generated keywords and toponyms for geo-referenced images. J. Am. Soc. Inf. Sci.. doi: 10.1002/asi.22738)
Abstract:
This article presents the results of a user evaluation of automatically generated concept keywords and place names (toponyms) for geo-referenced images. Automatically annotating images is becoming indispensable for effective information retrieval, since the number of geo-referenced images available online is growing, yet many images are insufficiently tagged or captioned to be efficiently searchable by standard information retrieval procedures. The Tripod project developed original methods for automatically annotating geo-referenced images by generating representations of the likely visible footprint of a geo-referenced image, and using this footprint to query spatial databases and web resources. These queries return raw lists of potential keywords and toponyms, which are subsequently filtered and ranked. This article reports on user experiments designed to evaluate the quality of the generated annotations. The experiments combined quantitative and qualitative approaches: To retrieve a large number of responses, participants rated the annotations in standardized online questionnaires that showed an image and its corresponding keywords. In addition, several focus groups provided rich qualitative information in open discussions. The results of the evaluation show that currently the annotation method performs better on rural images than on urban ones. Further, for each image at least one suitable keyword could be generated. The integration of heterogeneous data sources resulted in some images having a high level of noise in the form of obviously wrong or spurious keywords. The article discusses the evaluation itself and methods to improve the automatic generation of annotations.
An echo of Steve Newcomb’s semantic impedance appears at:
Despite many advances since Smeulders et al.’s (2002) classic paper that set out challenges in content-based image retrieval, the quality of both nonspecialist text-based and content-based image retrieval still appears to lag behind the quality of specialist text retrieval, and the semantic gap, identified by Smeulders et al. as a fundamental issue in content-based image retrieval, remains to be bridged. Smeulders defined the semantic gap as
the lack of coincidence between the information that one can extract from the visual data and the interpretation that the same data have for a user in a given situation. (p. 1353)
In fact, text-based systems that attempt to index images based on text thought to be relevant to an image, for example, by using image captions, tags, or text found near an image in a document, suffer from an identical problem. Since text is being used as a proxy by an individual in annotating image content, those querying a system may or may not have similar worldviews or conceptualizations as the annotator. (emphasis added)
That last sentence could have come out of a topic map book.
Curious what you make of the author’s claim that spatial locations provide an “external context” that bridges the “semantic gap?”
If we all use the same map of spatial locations, are you surprised by the lack of a “semantic gap?”
