Dense Non-rigid Structure from Motion

Abstract

This paper shows how current sparse approaches to Non-Rigid Structure From Motion (NRSfM) can be extended to template-free dense reconstruction. Our contribution can be broken down into two components: (i) We provide asymptotic improvements to current optimisation approaches, allowing them to scale to the dense case. (ii) We remove the need to estimate an initial rest shape or a known 3d template. This both increases the robustness of the method, allowing it to be applied to arbitrary sequences, and further decreases the run-time of our approach. Extending the existing techniques of NRSfM to fullyautomatic and dense template-free reconstruction substantially increases the difficulty of the optimisation problems faced, and we provide novel approaches for avoiding local minima in both discrete and continuous sub-problems.