Non-rigid Structure from Motion using Local Rigidity in Space & Time

Computer Vision (Dr. Sohaib Khan, Fall 2012) – Research Project – Later Extended & Published in IEEE WACV 2014

Collaborated with: Bilal Mahmood

Non-rigid structure from motion requires exploiting local rigidity between points (rigidity in space) and between adjacent frames (rigidity in time); focused to detect rigid points (or bone-lengths) without any prior knowledge of the data-set being input, making the solution more general.

Abstract: Reconstruction of non-rigid structure from motion is an ill-posed problem which requires some additional constraints to solve. Simple Tomasi-Kanade factorization approach does not apply to non-rigid structure from motion due to larger number of unknowns than constraints. As discussed recently in the unpublished paper by Ali Rehan [1], the structure can be assumed to be locally rigid in a small temporal window which allows for the use of Tomasi-Kanade’s method to produce stable results for that window. For utilizing rigidity in space, Rehan et al. used a provided list of bone length points to further constraint the problem.
In this paper we have explored several ways of adding length guidance to [1] without the use of any provided bone lengths. We will show that these methods produce comparable results to the case where bone length constraints are provided additionally along with the dataset.
Project Report
Paper (Published at IEEE WACV 2014)