Good features to track: A view geometric approach

Despite many alternatives to point tracking problem, iterative least squares solution solving the optical flow constraint has been the most popular approach used by many in the field. This paper attempts to leverage the former efforts to enhance such tracking methods by introducing geometric constraint to the tracking problem. In contrast to alternative geometry based methods, the proposed approach provides an online solution to optical flow estimation by exploiting Horn and Schunck flow estimation regularized by view geometric constraints. We particularly use the geometric invariants in the projective coordinates and conjecture that the traditional appearance solution needs to be geometrically regularized. Geometric regularization is achieved by estimating the posterior distribution of possible locations given the samples drawn from other tracked points which specify the scene geometry. Our experiments demonstrate that tracking can still be performed with great accuracy even when the features undergo appearance changes due to projective deformation of the template. Hence the resulting algorithm judges the quality of the tracked points based on not only appearance similarity but also geometric consistency.