Differential tracking with a kernel-based region covariance descriptor

Beyond Covariance: Feature Representation with Nonlinear Kernel Matrices

“…The second approach reduces the interference from the background within an image region. In object tracking [28], pixels are weighted in computing covariance matrix, and the farther a pixel is from the center of a region, the lower is its weight. The third approach, which may be most related to ours, considers to model high-order statistics of features [14,7].…”

Untitled

“…The second process of this technique makes the descriptor non-linear. Wu, et al [9] present a kernelbased region covariance descriptor for differential tracking.…”