For many cases, e.g. the alignment of timelapse images of a stable camera, the transformation between the images has less degrees of freedom than the full 2d homogeneous transformation matrix with translation vectors. It will for example not exhibit a coordinate system translation (if the camera position is truly fixed) and no scaling (in stable areas). Therefore, it might be better to restrict the transformation to simple rotations (which may however occur in the underlying 3d space and would be projected to the 2d image, --> look at the projective geometry of the underlying camera coordinate system?). This could improve robustness of alignment in these cases and should be implemented as an optional feature, as in other cases (hillshades from separate field campaigns), scaling and exact camera position may differ between the different images and the assumption of a full 2d homogeneous transformation matrix is justified.
For many cases, e.g. the alignment of timelapse images of a stable camera, the transformation between the images has less degrees of freedom than the full 2d homogeneous transformation matrix with translation vectors. It will for example not exhibit a coordinate system translation (if the camera position is truly fixed) and no scaling (in stable areas). Therefore, it might be better to restrict the transformation to simple rotations (which may however occur in the underlying 3d space and would be projected to the 2d image, --> look at the projective geometry of the underlying camera coordinate system?). This could improve robustness of alignment in these cases and should be implemented as an optional feature, as in other cases (hillshades from separate field campaigns), scaling and exact camera position may differ between the different images and the assumption of a full 2d homogeneous transformation matrix is justified.