Motion Annotation Programs: A Scalable Approach to Annotating Kinematic. Articulations in Large 3D Shape Collections

Abstract:

3D models of real-world objects are essential for many applications, including the creation of virtual environments for AI training. To mimic real-world objects in these applications, objects must be annotated with their kinematic mobilities. Annotating kinematic motions is time-consuming, and it is not well-suited to typical crowdsourcing workflows due to the significant domain expertise required. In this paper, we present a system that helps individual expert users rapidly annotate kinematic motions in large 3D shape collections. The organizing concept of our system is motion annotation programs: simple, re-usable procedural rules that generate motion for a given input shape. Our interactive system allows users to author these rules and quickly apply them to collections of functionally-related objects. Using our system, an expert annotated over 1000 joints in under 3 hours. In a user study, participants with no prior experience with our system were able to annotate motions 1.5x faster than with a baseline manual annotation tool.