Computing the Cost of Occlusion

Recently, Cox et al. (1996, CVGIP: Image Understanding63, 542–567) presented a new dynamic programming-based stereo matching algorithm. The algorithm uses a parameter which represents the cost of occlusion. This cost is levied if the algorithm decides that two measurements, each from a different camera along corresponding epipolar lines, are not projections of the same point in space. The occlusion cost is dependent on the standard deviation of the (Gaussian) sensor noise, σ, and the probability of match detection, PD. Under certain conditions such as low signal-to-noise ratio, the algorithm of Cox et al. will declare occlusions where they do not exist. We offer an alternative definition for the cost of occlusion, based on a decision-theoretic formulation for the matching process. This alternative improves the performance of the matching algorithm.