but they still need interpolation because pixels in general won't exactly overlap.
as far as i can see it's replacing the interpolation in 2d with multiple interpolations in 1d. but perhaps i am missing something.
[edit] ok, after skimming the paper, it seems that they approximate the interpolation (using nearest neighbour). which makes it very fast. so really it's a neat way of working out what nearest neighbour is in 2d by doing it in 1d multiple times.
Note that one could also do "area mapping" with nearest neighbor instead of 4-pixel interpolation (if speed is more important than smoothness) to avoid the white dots. This is the most obvious "naive" approach to me -- basically the same cost as the white dots approach (loop over every destination pixel just once, grab just one pixel from source) but with some repeated neighboring pixels instead of missing information.
as far as i can see it's replacing the interpolation in 2d with multiple interpolations in 1d. but perhaps i am missing something.
[edit] ok, after skimming the paper, it seems that they approximate the interpolation (using nearest neighbour). which makes it very fast. so really it's a neat way of working out what nearest neighbour is in 2d by doing it in 1d multiple times.