Robust Image Matching under Partial Occlusion and Spatially Varying Illumination Change

Robust and Efficient Image Alignment Based on Relative Gradient Matching

“…Note that the weight is determined from the image grayscale to assign small weights for very dark or very bright grayscales. The selection of the weight function is similar to that described in [9]. This is used to alleviate the problem of relative gradient matching in the shadow or brightness saturation regions.…”

“…This energy minimization is a nonlinear least square minimization problem. When a good initial guess is available, we can employ Newton method to solve this minimization problem very efficiently [9], [10]. In general, when an initial guess for the transformation parameters is given as , we can update the transformation parameters by minimizing the following energy function: (8) Note that the inverse transformation can be proved to be given by (9) with , where is a rotation matrix with rotation angle .…”

“…When a good initial guess is available, we can employ Newton method to solve this minimization problem very efficiently [9], [10]. In general, when an initial guess for the transformation parameters is given as , we can update the transformation parameters by minimizing the following energy function: (8) Note that the inverse transformation can be proved to be given by (9) with , where is a rotation matrix with rotation angle . When applying Newton method to minimize the dynamic twice-differentiable energy function in (8) with respect to the transformation parameters, we need to compute the function values , , and their partial derivatives at subpixel image coordinates.…”

Class-Based Matching of Object Parts

“…Typically, affine or low-order polynomial models of illumination and geometry are used (e.g. [11]), but more complicated schemes [5,12] have also been proposed. If estimates of the parameters are known, the features may be matched across transformation (e.g.…”

“…The drawback is the necessity to use video sequences, which are not always available, and require additional effort to capture, store and manipulate. In addition, the approximation provided by common parametric models [11] is only valid for a limited range of transformations.…”