Given a 3D shape, how do you retrieve similar shapes from a 3D shape database, or from the a range of storage areas on the web? This is the 3D shape retrieval problem and it has many applications, such as finding machine parts, retrieving shapes from a repository of CAD models, matching shapes across a set of museum artefacts and so on. One of the central problems is to build a system that can cope with translation, rotation and scale variations, while retaining the ability to finely discriminate between object classes. This project will explore a range of techniques available in the literature, starting with the Light Field Descriptor method proposed by Chen et al  . In this method, an object is characterised by its visual appearance from multiple angles distributed across a viewing sphere. The method has given very high performance on the Princeton Shape Benchmark , but this comes at a high computational cost. We will explore ideas for both improving classification performance and increasing shape retrieval efficiency. You will implement a system in MATLAB and compare the performance of a range of 3D shape retrieval system variants.
PAT: highly desirable; CVI: desirable
(1) Chen, Tian and Ouhyoung, On visual similarity based 3D model retrieval, Eurographics 2003.
(2) Shilane, Min, Kazhdan and Funkhouser, The Princeton Shape Benchmark, Shape Modelling International, Genova , Italy, 2004