Model-Based Motion Estimation for Synthetic Images
Maneesh Agrawala, Andrew Beers, Navin Chaddha
Abstract
One approach to performing motion estimation on synthetic animations is to treat them as video sequences and use standard image-based motion estimation methods. Alternatively, we can take advantage of information used in rendering the animation to guide the motion estimation algorithm. This information includes the 3D movements of the objects in the scene and the projection transformations from 3D world space into screen space. In this paper we examine how to use this high level object motion information to perform fast, accurate block-based motion estimation for synthetic animations.
The optical flow field is a 2D vector field describing the translational motion of each pixel from frame to frame. Our motion estimation algorithm first computes the optical flow field, based on the object motion information. We then combine the per-pixel motion information for a block of pixels to create a single 2D projective matrix that best encodes the motion of all the pixels in the block. The entries of the 2D matrix are determined using a least squares formulation. Our algorithms are more accurate and much faster in algorithmic complexity than most image-based motion estimation algorithms.