3-D content-based retrieval and classification with applications to museum data
There is an increasing number of multimedia collections arising in areas once only the domain of text and 2-D images. Richer types of multimedia such as audio, video and 3-D objects are becoming more and more common place. However, current retrieval techniques in these areas are not as sophisticated as textual and 2-D image techniques and in many cases rely upon textual searching through associated keywords. This thesis is concerned with the retrieval of 3-D objects and with the application of these techniques to the problem of 3-D object annotation. The majority of the work in this thesis has been driven by the European project, SCULPTEUR. This thesis provides an in-depth analysis of a range of 3-D shape descriptors for their suitability for general purpose and specific retrieval tasks using a publicly available data set, the Princeton Shape Benchmark, and using real world museum objects evaluated using a variety of performance metrics. This thesis also investigates the use of 3-D shape descriptors as inputs to popular classification algorithms and a novel classifier agent for use with the SCULPTEUR system is designed and developed and its performance analysed. Several techniques are investigated to improve individual classifier performance. One set of techniques combines several classifiers whereas the other set of techniques aim to find the optimal training parameters for a classifier. The final chapter of this thesis explores a possible application of these techniques to the problem of 3-D object annotation.