神经网络来预测两个图像块的匹配程度,使用它来计算立体匹配代价。代价定义为基于交叉的代价聚合和半全局匹配,利用左右一致性检查来消除在遮挡区域的误差。 [18] MatchNet: Unifying...立体匹配共四步:匹配代价计算、匹配代价聚集、视差计算、视差优化立体匹配主要解决两个问题:一个是低纹理区域的视差估计一个是深度不...
Specially for leftmost and rightmost regions, the segment-based refinement can greatly improve the mismatched disparity holes. The experimental results show that the proposed system can achieve higher accurate depth maps than the conventional methods....
Similarly, we replicate the plane refinement in area 𝑀𝑝Mp. The plane refinement assigns two separate label spaces for 𝑈𝑝Up and 𝑀𝑝Mp. Lastly, the proposal by RANSAC fitting is employed to the center grid 𝐶𝑖𝑗Cij, which can estimate the 3D label of a disparity plane ...