The effect of spatial resolution reduction techniques on the temporal properties of video sequences

Abstract

Singular value decomposition (SVD) is a common technique that is performed on video sequences in a number of computer vision and robotics applications. The left singular vectors represent the eigenimages, while the right singular vectors represent the temporal properties of the video sequence. It is obvious that spatial reduction techniques affect the left singular vectors, however, the extent of their effect on the right singular vectors is not clear. Understanding how the right singular vectors are affected is important because many SVD algorithms rely on computing them as an intermediate step to computing the eigenimages. The work presented here quantifies the effects of different spatial resolution reduction techniques on the right singular vectors that are computed from those video sequences. Examples show that using random sampling for spatial resolution reduction rather than a low-pass filtering technique results in less perturbation of the temporal properties.