Speaker: Kurt Leimer
With online 3D-model repositories growing ever larger, modelers are presented with new opportunities to draw inspiration from existing shapes or even create new models by combining different parts of these shapes. However, efficiently browsing these vast collections can be a difficult task, especially since they often lack a meaningful labeling of shapes or shape parts. In this thesis we will propose a new parameterization of 3D shape collections that provides the user with an intuitive way to browse collections of models. Given a set of 3D shapes from the same family, we first obtain a consistent co-segmentation by performing co-analysis. This is done by first segmenting each shape individually and then clustering the segments across all shapes based on a number of face-level features, thus obtaining a labeling of shape parts. Next, an abstract representation of each shape is created by computing graphs in which nodes correspond to shape parts and edges correspond to relations between shape parts, such as adjacency or symmetry. These relations, combined with additional values such as the distance or the difference in scale between parts, serve to create a new parameterization of the shape collection. With this parameterization the user is able to explore the shape collection by starting with an existing shape and altering the parameters of the shape parts, either freely or within a lower dimensional subspace based on the parameters with the largest variations across a part category.