Motion Estimation for Large Displacements and Deformations

Motion Estimation for Large Displacements and Deformations
Reconstructed pointcloud of Montreal, QC, Canada. The dataset was provided by Valcartier Defence Research and Development Canada (DRDC).

The preprint of our paper Motion Estimation for Large Displacements and Deformations -published in Scientific Reports- is available on arxiv.org. The work is co-authored by Qiao Chen and Charalambos Poullis. The camera-ready paper is now available: https://www.nature.com/articles/s41598-022-21987-7

The work was featured in LeDevoir (article).

TL;DR: Large-scale images such as wide-area motion imagery cannot be handled by deep learning architectures. The standard procedural method for image-based reconstruction is Structure-from-Motion followed by Multiview Stereo which is time-consuming and can result in incomplete and disjoint reconstructions. In this paper, we propose a variational optical flow method based on graph and feature matching tailored for the dense matching and reconstruction of large-scale images. The proposed method surpasses all variational methods and gives comparable results with the latest deep learning approaches.