Out-lier removal algorithm for model-based coded video

Abstract

Model-based coding extracts texture from images of objects and projects them onto computer models of the same objects. Low bit rate transmission is achieved by approximating the shape and movement of the objects and relaying the parameters to the decoder. To estimate the motion of the models a number of features must be accurately tracked inter-frame. The proposed algorithm employs a large displacement vector set, and then proceeds to iteratively remove points exhibiting large disagreement between the 2D translational field and a projection of the estimated 3D rotation. Subjective improvement in motion estimation and better quantitative agreement between motion fields are reported.