This paper presents a method that generates a mosaic image from multiple images or a sequence of images without any assumptions on camera intrinsic parameters or camera motion. The most appropriate transform model is automatically selected according to the global alignment error. Given a set of images, the optimal parameters for a 2D projective transform model are computed using a linear least squares estimation method and then iterative nonlinear optimization method. When the error exceeds a predefined limit, a more simple transform model is applied to the alignment. Our experiments showed that false alignments are reduced even in the case of 360° panoramic images.