Consolidating Segmentwise Non-Rigid Structure from Motion

Vladislav Golyanik, André Jonas, Didier Stricker

In: International Conference on Machine Vision Applications (MVA). IAPR Conference on Machine Vision Applications (MVA-2019) May 27-31 Tokyo Japan IAPR 2019.


This paper introduces a new segmentwise technique which consolidates multiple principles for non-rigid structure from motion (NRSfM) into a single energy-based framework. The energy functional of our Consolidating Monocular Dynamic Reconstruction (CMDR) approach is optimised by non-linear least squares and includes terms allowing to define the deformation model and additional constraints simultaneously in the metric and trajectory spaces. The proposed method achieves high accuracy on several tested sequences while providing robustness and scalability due to the spatial scene segmentation and the new lifted spatial Laplacian term. CMDR is flexible and easy to implement, thanks to the unified optimisation framework. It allows for scenario-specific extensions and can be used for rapid prototyping of new NRSfM methods.



Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence