Frankenhorse: Automatic Completion of Articulating Objects from Image-based Reconstruction
In Proceedings British Machine Vision Conference 2014
AbstractStructure from Motion reconstructions of objects often contain holes. While small holes can be completed with a smoothness prior, completing large holes requires a higher-level understanding of the object. We present a method to complete large holes in articulating objects by reconstructing and aligning sets of objects of the same class, using the well-reconstructed parts in each model to complete holes in the others, resulting in a `Frankenhorse' completion. Our proposed method is fully automatic, yet is able to handle articulation, intra-class variation, holes and clutter present in the reconstructions. This is achieved through our novel segmentation and clutter removal processes as well as by the use of a robust method for piecewise-rigid registration of the models. We show that our method can fill large holes even when only a small set of models with high variability and low reconstruction quality is available.
FilesExtended Abstract (PDF, 1 page, 3.8M)
Paper (PDF, 12 pages, 8.6M)
Supplemental Materials (ZIP, 9.4M)