3D Shape Analysis

123D Shape Retrieval

12.1 Introduction

The need for efficient 3D search engines is mainly motivated by the large number of 3D models that are being created by novice and professional users, used and stored in public and in private domains, and shared at large scale. In a nutshell, a 3D search engine has the same structure as any image and text search engine. In an offline step, all the 3D models in a collection are indexed by computing, for each 3D model, a set of numerical descriptors which represent uniquely their shape. This is the feature extraction stage. (For details about 3D shape description methods, we refer the reader to Chapters and 5.) The 3D models that are similar in shape should have similar descriptors, while the 3D models that differ in shape should have different descriptors. At runtime, the user specifies a query. The system then ranks the models in the collection according to the distances computed between their descriptors and the descriptor(s) of the query and returns the top elements in the ranked list. This is the similarity computation stage.

In general, the query can have different forms; the user can, for instance, type a set of keywords, draw (a set of) 2D sketch(es), use a few images, or even use another 3D model. We refer to the first three cases as cross‐domain retrieval since the query lies in a domain that is different from the domain of the indexed 3D models. This will be covered in Chapter 13. In this chapter, we focus on the tools and techniques ...