Support-driven Shape Analysis

Description

3D models are nowadays of high demand especially by the applications such as virtualreality and 3D printing. However, 3D modeling from scratch is rather difficult forunskilled users and still time-consuming even for professionals. On the other hand, theavailability of large collections of 3D models together with various shape segmentationapproaches offers a great opportunity for composing new 3D models by recombining theexisting model parts. The key challenge is how to properly encode relationships betweenparts in existing models by shape analysis so that novel but still functionally plausiblemodel variations can be easily created.Existing solutions often regard symmetry as the main semantic cue for encoding part-levelrelationships. However, such approaches are not applicable to objects or scenesexhibiting little or no symmetry. On the other hand, support with stability is anotherfundamental relationship between parts of objects in the physical world, especially forman-made objects. This motivates us to perform shape analysis in a support-drivenmanner, where the symmetry cue if applicable is still useful but becomes supplementary.In this project we will focus on shape analysis of man-made objects and explore a novelsupport-driven structural organization of their object parts. We are interested indefining and detecting special subsets of object parts with support and stability inexisting models. We call such special subsets as support substructures. Using suchsupport substructures instead of semantic parts as building blocks, makes it possible tosynthesize both in-class and across-class variations. One of the key challenges in thiswork is to find a relatively small set of support substructures for easy creation of manynovel, non-trivial, and interesting shape variations.We will show that such high-level shape analysis and the resulting supportsubstructures benefit various applications, e.g., part-based shape reshuffling betweenmodels, structure rearrangement and synthesis for a single input model. We believe thatour work will turn the attention of the research community more towards support andstability, which are very useful but have been largely unexplored for shape analysis andsynthesis.