Semantic-oriented 3d shape retrieval using relevance feedback

Abstract

Shape-based retrieval of 3D models has become an important challenge in computer graphics. Object similarity, however, is a subjective matter, dependent on the human viewer, since objects have semantics and are not mere geometric entities. Relevance feedback aims at addressing the subjectivity of similarity. This paper presents a novel relevance feedback algorithm that is based on supervised as well as unsupervised feature extraction techniques. It also proposes a novel signature for 3D models, the sphere projection. A Web search engine that realizes the signature and the relevance feedback algorithm, is presented. We show that the proposed approach produces good results and outperforms previous techniques.

Comparison of Dimension Reduction Methods for Database-Adaptive 3D Model Retrieval

“…Interactive relevance feedback, a form of on-line interactive learning, has been explored by several re-searchers for 3D model retrieval [8,1,15,16]. Elad et al is among the first to apply Support Vector Machines (SVM) learning in an on-line learning setting to improve 3D model retrieval [8].…”

“…Elad et al is among the first to apply Support Vector Machines (SVM) learning in an on-line learning setting to improve 3D model retrieval [8]. Leifman et al [15] performed Kernel Principal Component Analysis (Kernel PCA) [10] for an unsupervised learning of a feature subspace before applying a relevance feedback technique that employs Biased Discriminant Analysis (BDA) or Linear Discriminant Analysis (LDA) on the learned subspace. Novotni et al [16] compared several learning methods, SVM, BDA, and Kernel-BDA, for their retrieval performance in a relevance feedback setting.…”

“…Unlike relevance feedback, unsupervised off-line learning has seen very little attention in 3D model retrieval. The Kernel-PCA employed by Leifman et al [15] is an example. The purity proposed by Bustos et al [5] can also be considered as a weak form of unsupervised off-line learning.…”

Shape-Based Autotagging of 3D Models for Retrieval

“…To retrieve 3D models, a query for the Query-ByExample (QBE) approach could be sketches in either 2D [12] or 3D of the desired shape, an example 3D model, or a photograph of a real-world object having the desired shape. Most of the published work on 3D model retrieval used the QBE approach [2,5,6,7,9,10,12,14]. While the QBE approach is quite powerful and useful in many application scenarios, it does have drawbacks.…”

“…Capturing semantic aspect of the desired shape is also difficult with the QBE approach. Previous work used either single-class learning via relevance feedback [6,7] or off-line multi-class learning [10] for semantics.…”