Seam maps

Imagine pulling all of the seams out of an image in a particular direction. You label all the pixels belonging to the first seam removed “1”, all the pixels belonging to the second seam removed “2”, and so on. Once you have pulled all the seams out, you have a seam map[1] in which every pixel is labelled with the number of the seam it belongs to.

Now an image with $m$ seams removed can be reconstructed by gathering together only pixels from the original image whose label in the seam map is greater than $m$ . This is how “real-time” seam carving is performed (the precomputation of the map only takes a few seconds).

Seamstress provides an easy interface to automatically compute seam maps for you (see Section 4.4).