Robust Multi-homography Method for Image Stitching under Large Viewpoint Changes

Abstract

Image stitching technique has many applications in computer vision. Traditional approaches have achieved much success under the assumption that the input images must have little or no parallax. However, this assumption cannot be guaranteed in real-world problems, for example, building images with ground under large viewpoint changes. The traditional methods only use a whole homography for alignment, which will result in severe distortions. To deal with the problem of stitching building images with ground under large viewpoint changes, we propose a robust multi-homography image composition method. It mixes multiple homographies for stitching building images with ground under large viewpoint changes. By calculating different homographies from different types of features, multiple homographies are then be blended with Gaussian weights to construct panorama. The chosen images under various conditions are used to evaluate our proposed method with other methods. The experimental results demonstrate that our proposed method performs higher alignment accuracy and more natural panorama than the traditional projective transformation and some representative state-of-the-art image stitching methods.