Differential tracking through sampling and linearizing the local appearance manifold

Recovering motion information from input camera image sequences is a classic problem of computer vision. Conventional approaches estimate motion from either dense optical flow or sparse feature correspondences identified across successive image frames. Among other things, performance depends on the accuracy of the feature detection, which can be problematic in scenes that exhibit view-dependent geometric or photometric behaviors such as occlusion, semi-transparancy, specularity and curved reflections. Beyond feature measurements, researchers have also developed approaches that directly utilize appearance (intensity) measurements. Such appearance-based approaches eliminate the need for feature extraction and avoid the difficulty of identifying correspondences. However the simplicity of on-line processing of image features is usually traded for complexity in off-line modeling of the appearance function. Because the appearance function is typically very nonlinear, learning it usually requires an impractically large number of training samples.;I will present a novel appearance-based framework that can be used to estimate rigid motion in a manner that is computationally simple and does not require global modeling of the appearance function. The basic idea is as follows. An n-pixel image can be considered as a point in an n-dimensional appearance space. When an object in the scene or the camera moves, the image point moves along a low-dimensional appearance manifold. While globally nonlinear, the appearance manifold can be locally linearized using a small number of nearby image samples. This linear approximation of the local appearance manifold defines a mapping between the images and the underlying motion parameters, allowing the motion estimation to be formulated as solving a linear system.;I will address three key issues related to motion estimation: how to acquire local appearance samples, how to derive a local linear approximation given appearance samples, and whether the linear approximation is sufficiently close to the real local appearance manifold. In addition I will present a novel approach to motion segmentation that utilizes the same appearance-based framework to classify individual image pixels into groups associated with different underlying rigid motions.