Recall that DMAG is based on optical flow (variational method), DMAG2 relies on window/block matching (local method), and DMAG3 is based on graph cuts (global method).
I should really consider a bunch of stereo pairs to do this study but, since I am not really a scholar but merely an hobbyist, a single stereo pair will suffice. Here's the stereo pair I am gonna use to compare the three depth map automatic generators:
The stereo pair above was taken in Boston's North End with my trusted Fuji W3 (when it doesn't have focusing issues). Important: it was aligned in StereoPhoto Maker using the automatic alignment tool.
DMAG finds the optical flow (displacement) between the two images of a stereo pair, which is equivalent to finding the disparity between the two images. This is a variational method which solves for a continuous displacement.
I have generated three depths maps using different values for the smoothness parameter, the most important parameter:
Recall that these depth maps are actually optical flow displacement maps (black is close and white is far), in other words, inverted depth maps (where white is close and black is far). I am gonna stick with the default (smoothness parameter = 18) to generate the intermediate synthesized views (five frames) with FSG. Here's the animated 3d gif:
DMAG2 is a disparity map generator which, for every pixel in the left image, finds the best matching pixel in the right image using windows/blocks centered on the pixels of interest. It's a purely local method that uses the adaptive window concept via the use of window weights. As with other local methods, there's no smoothing of the results, which can be hard on the eyes.
Using the results of DMAG (min and max displacements) and the trusted DF2 (Disparity Finder 2), I have come to the conclusion that the min disparity is -20 (furthest background object) and the max disparity is 0 (closest foreground object). These are the values I am gonna use in DMAG2 (and DMAG3).
Because the thing is rather slow (due to window weight computation), I am just gonna use the default value for the Window Radius (12), which is the most important parameter. This is the depth map that was obtained:
Here's the animated 3d gif obtained with FSG2, again using five frames:
DMAG3 is a so-called global method based on the Graph Cut methodology. It's probably the most popular way to find the stereo correspondence between the two images of a stereo pair.
The first thing I did was to generate four depth maps playing with the smoothness parameter and the occlusion penalty:
I am gonna keep the depth map obtained with smoothness parameter = 10 (default) and occlusion penalty = 30 and generate a 3d animated gif using FSG2 for the intermediate frames:
Well, after all this, I am not gonna draw any conclusions (bummer) and let the reader decide which is best. Note that the stereo pair you start with as well as the choice of parameters play a big role in the results you're gonna get. I don't think there's an automatic depth map generator that consistently gets better results than others.