Learning to predict part mobility from a single static snapshot

摘要: We introduce a method for learning model the mobility of parts in 3D objects. Our allows not only to understand dynamic functionalities one or more object, but also apply functions static models. Specifically, learned part can predict mobilities object given form single snapshot reflecting spatial configuration space, and transfer from relevant units training data. The data consists set different motion types. Each unit is composed pair (one moving reference part), along with usage examples consisting few snapshots capturing states unit. Taking advantage linearity characteristic exhibited by most motions everyday objects, utilizing part-relation descriptors, we define mapping units. This employs motion-dependent snapshot-to-unit distance obtained via metric learning. show that our scheme leads accurate prediction proper transfer. demonstrate other applications such as motion-driven detection hierarchy construction.