Content-Based Image Retrieval is important for two reasons. First, the oft-cited growth of image archives in many fields, and the rapid expansion of the Web, mean that successful image retrieval systems are fast becoming a necessity if the mass of accumulated data is to be useful. Second, database retrieval provides a framework within which the important questions of machine vision are brought into focus: successful retrieval is likely to require genuine image understanding. In view of these points, the evaluatio- n of retrieval systems becomes a matter of priority. There is already a substantial literature evaluating specific systems, but little high-level discussion of the evaluation methodologies themselves seems to have taken place. In the first part of the report, we propose a framework within which such issues can be addressed, analyse possible evaluation methodologies, indicate where they are appropriate and where they are not, and critique query-by-example and evaluation methodologies related to it. In the second part of the report, we apply the results of this analysis to a particular dataset. The dataset is problematic but typical: no ground truth is available for its semantics. Considering retrieval based on image segmentation- s, we present a novel method for its evaluation. Unlike methods of evaluation that rely on the existence or creation of ground truth, the proposed evaluatio- n procedure subjects human subjects to a psychovisual test comparing the results of different segmentation schemes. The test is designed to answer two questions: does consensus about a `best' segmentation exist, and if it does, what do we learn about segmentation schemes for retrieval? The results confirm that human subjects are consistent in their judgements, thus allowing meaningful evaluation.