Comparison of the old and new version of GCS

GCS 1.0

GCS 2.0 (new version)

Figure 1. Resulting color-coded disparity maps.

Note, the disparity map is denser with the new version of GCS. Small gaps between disparity levels seen as gray contour regions in the old version has almost disappeared in the new version. These artifacts were removed due to a better occlusion model.

The old occlusion model was the uniqueness, see the inhibition zone in Fig. 2(a), or in the original paper [1, Fig. 2(b)]. The final robust matching [1, p. 4] tends to reject both competing disparity hypotheses when they are around modulo 0.5 pixel. Such competing hypotheses have similar (rather low) correlation, and according to the stability principle, none of them become a part of the solution.

(a) Inhibition zone of the old version.

(b) The new inhibition zone. It has a gap.

Figure 2. The inhibition zones.

Therefore, we redefine the occlusion model. This is implemented in GCS version 2.0. The new inhibition zone is shown in Fig. 2(b). The inhibition zone gap changes the mutual exclusivity in the correspondence hypothesis set. The candidates with 0.5 pixel disparity are not competitors any more. The drawback is that the uniqueness does not hold. To obtain one-to-one matching, we use a simple procedure, which aggregates pixels violating the uniqueness and computes the final disparity as a correlation weighted average. This even produces a rough sub-pixel estimate.