Generalized Dynamic 3D Reconstruction

Image-based dynamic reconstruction models estimate the geometry (i.e. the Euclidean coordinates) of temporally evolving 3D points from 2D feature observations, striving to enrich the understanding and visualization of events captured by uncontrolled and heterogeneous imagery. Existing works on dynamic 3D reconstruction constrain geometry through ad-hoc priors or supervision through domain-specific datasets. We explore transformative graph-theoretic geometric formulations able to redefine the state of the art in robustness and accuracy. We also develop learningbased methods taking advantage of 3D deep network architecture to improve computational efficiency and reconstruction accuracy without losing generality. In doing so, we develop integrative frameworks combining rigorous geometric-based formulations along with data-driven 3D motion semantics. We also propose an framework that jointly learns feature detection, descriptor representation, and cross-frame matching based on weak supervision for relative camera pose estimation task which is also a fundamental building block for the dynamic reconstruction problem.