To show the robustness of our method, the pictures in this stack were captured with a predefined delay (two 33ms frames were dropped between the capture of the source and reference images). Note that the reference image contains extremely dark regions. Moreover, large parts of the scene are relatively flat, making it even more difficult to find correspondences. Here we warp the source image using our method ('3'), replacing our matcher with SIFT ('4'), and using the method by Bao et al. ('5').