Research And System Implementation Of Confidence Propagation Based Patch-Match Multi-View Depth Estimation

Albeit PatchMatch based multi-view stereo has made remarkable progress to improve the accuracy and completeness of 3D reconstruction,depth estimation in low-textured areas is still a challenging problem due to the matching ambiguity,which leads to the lack of reconstructed point cloud on the low-textured surface of the scene.In this thsis,we deeply investigated the PatchMatch multi-view depth estimation algorithm,and propose a novel method,confidence propagation based PatchMatch multi-view depth estimation algorithm to solve this issue.The main tasks are as follows:(1)The concept of matching confidence is defined in multi-view stereo matching.Based on the variance of the reference image patch,a pixel-wise coefficient is calculated to quantify the reliability of matching measure.(2)A confidence-aware adaptive propagation strategy is proposed.Based on the matching confidence,additional sampling pixels are selected to expands the hypotheses set.At the same time,the matching measure is modified to propagate the reliable candidate hypotheses to the low-confidence pixels.(3)A cost-radio based confidence update strategy is proposed.Based on the random search step of the original PatchMatch algorithm,this paper assusme that the radio of the photometric cost before and after the disturbance reflects the effectiveness of matching on current pixel.The matching confidence is updated by the cost-radio.(4)A depth estimation algorithm of large-scale untextured regions is proposed.In order to solve the problem that pixels in these regions quickly converge to anomaly estimation,a forced update sheme is proposed to start from the edge pixels.A binary smoothness term is also proposed to constraint the estimated difference of adjacent pixels.Experimental results demonstrate the algorithm proposed in this paper significantly improves the completeness of the reconstructed model,especially for low-textured scenes.The multi-view depth estimation system designed by this algorithm can compute high-precision depth maps and high-quality 3D models.