Occlusions and visibility

In a triangulation sensor with two viewpoints $k$ and $l$ two types of occlusion occur. If parts of the object are hidden in both viewpoints due to object self-occlusion, then we speak of object occlusions which cannot be resolved from this viewpoint. If a surface region is visible in viewpoint $k$ but not in $l$, we speak of a shadow occlusion. The regions have a shadow-like appearance of undefined disparity values since the occlusions at view $l$ cast a shadow on the object as seen from view $k$. Shadow occlusions are in fact detected by the uniqueness constraint discussed in section 7.1. A solution to avoid shadow occlusions is to incorporate a symmetrical multi-viewpoint matcher as proposed in this contribution. Points that are shadowed in the (right) view $k+1$ are normally visible in the (left) view $k-1$ and vice versa. The exploitation of up-and down-links will resolve for most of the shadow occlusions. A helpful measure in this context is the visibility V that defines for a pixel in view $k$ the maximum number of possible correspondences in the sequence. $V = 1$ is caused by a shadow occlusion, $V >= 2$ allows a depth estimate.