Abstract

Many non-rigid 3D structures are not modelled wellthrough a low-rank subspace assumption. This is problem-atic when it comes to their reconstruction through Structurefrom Motion (SfM). We argue in this paper that a more ex-pressive and general assumption can be made around com-pressible 3D structures. The vision community, however,has hitherto struggled to formulate effective strategies forrecovering such structures after projection without the aidof additional priors (e.g. temporal ordering, rigid substruc-tures, etc.). In this paper we present a “prior-less” ap-proach to solve compressible SfM. Specifically, we demon-strate how the problem of SfM assuming compressible 3Dstructures can be theoretically characterized as a blocksparse dictionary learning problem. We validate our ap-proach experimentally by demonstrating reconstructions of3D structures that are intractable using current state-of-the-art low-rank SfM approaches.