Use of the Internet for retrieving images has gained considerable momentum. Searching the Internet for images based on content is an important task, not only in engineering but also for regular Internet users. Due to technological developments, users want their search results to be more accurate and returned faster. This presents many challenges--such as multi-dimensional keys and ways to reduce the number of nodes to be traversed during the query--for researchers working in content-based image retrieval (CBIR).
Popular data structures for CBIR include the kd tree and the nearest neighbor (NN) search method. This paper provides a framework for assessing the performance of kd trees and NN searches using a new cost model. The authors address the major issues in kd trees before presenting their approach. They consider two of the commonly used traversal approaches for this evaluation: depth-first traversal and best-bin-first search. Evaluation of these approaches is done with and without culling in terms of average traversal node and distance error ratio. People working in CBIR will find this well-written paper useful.