Data-driven approach for automatic orientation of 3D shapes

Visualization and visual browsing of 3D model collections require rendering the 3D models from viewpoints that allow the viewer to distinguish between them. In this paper, we introduce a new framework for the automatic selection of the best views of 3D models. We build on the assumption that models belonging to the same class of shapes share the same salient features that discriminate them from the models of other classes. This allows us to formulate the best-view selection problem as a feature selection and classification task. First a 3D model is described with a set of view-based descriptors characterizing the appearance of the model when it is seen from different viewpoints. In a second step we train a classifier that learns for each shape class the set of 2D views that maximize the intra-class similarity and the inter-class dissimilarities. Finally, we post-process the selected 2D views to estimate their upright orientation. We exploit the fact that most of natural and man-made shapes are symmetric and their upright orientation is aligned with their major axis of symmetry. Experiments on the best-view selection benchmark demonstrate that the estimated best views with our data-driven approach are robust to intra-class variations and are consistent within the models of the same class of shapes. This makes the approach suitable for online visual browsing of large 3D data collections.