OBER-cross+ANP
Occluded Bilateral EPI Regularization + adaptive filter strength + normal + plane regularization
The method is based on a regularization scheme which consists of a bilateral filter as smoothing term, and an error metric which evaluates the variance in the image domain for a given 3D point.
The error metric includes an explicit occlusion term which handles single and double occlusions. Due to hard cutoff thresholds in the bilateral smoothing and the occlusion term this yields very fine details and sharp occlusion borders. The optimization is performed using trial and error evaluation of random disparity estimates and propagation of neighboring disparities. Due to the highly non-smooth and non-differential nature of the minimization term, convergence is problematic and requires careful initialization. The initialization therefore uses the results of the zero-crossing (also used by ZCTV) method as a starting point for the optimization.
This is the cross-setup version, using 17 from 81 views.
This version extends the original OBER with an adaptive filter strength + normal and plane regularization.
cross
C++
GNU/Linux 4.10.11 , i7-2600K @ 3.4Ghz
