Supervised Learning of Salient 2D Views of 3D Models

Abstract

We introduce a new framework for the automatic selection of the bestviews of 3D models based on the assumption that models belongingto the same class ofshapes share the same salient features. The main issue is learningthese features. We propose an algorithm for computing these featuresand their corresponding saliencyvalue. At the learning stage, a large set of features are computedfrom every model and a boosting algorithm is applied to learn theclassification function in the featurespace. AdaBoost learns a classifier that relies on a small subsetof the features with the mean of weak classifiers, and provides anefficient way for feature selectionand combination. Moreover it assigns weights to the selected featureswhich we interpret as a measure of the feature saliency within theclass. Our experiments usingthe LightField (LFD) descriptors and the Princeton Shape Benchmarkshow the suitability of the approach to 3D shape classification andbest-view selection for onlinevisual browsing of 3D data collections