Possibly misleading/incorrect bibliography entry (Feature #1627)

The source for the cost function used by StereoSGBM is given as Birchfield & Tomasi (1996) "Depth Discontinuities by Pixel-to-Pixel Stereo." This is the same source suggested in Hirshmuller (2008) as an alternative to using Mutual Information.

While the Birchfield & Tomasi (1996) paper does describe sampling insensitive sub-pixel dissimilarity, the primary purpose of the paper is to introduce a more specific cost function this is specific for detecting depth discontinuities. This cost function includes additional terms that penalize discontinuities in depth, making it less sensitive to small depth variations within a object.

I believe the cost function that is actually implemented in StereoSGBM does not include these additional discontinuity terms, so it is misleading to cite the 1996 paper. A more correct reference would be Birchfield & Tomasi (1998) "A Pixel Dissimilarity Measure that is Insensitive to Image Sampling," which describes the simpler cost function in detail.

If StereoSGBM does in fact use the cost function that is specialized for depth discontinuities, it would be good for this to be noted in the documentation.