基于多尺度注意力网络的立体匹配方法

Stereo Matching Method Based on Multiscale Attention Network

为解决网络深度与训练图像块大小耦合问题及进一步提高弱纹理区域及边缘处的匹配精度,提出了一种基于多尺度注意力网络的立体匹配方法.该方法将立体匹配过程分为2个阶段:第1阶段提出了一种成本网络用于计算匹配成本,该网络由基础网络层和缩放层组成.第2阶段提出了一种基于多尺度注意力的视差求精网络,该视差求精网络综合了多种视差线索,并加入多尺度注意力机制进一步提高立体匹配精度.该方法在KITTI 2012、KITTI 2015和SceneFlow数据集上的3像素坏点百分比分别为1.13%,1.87%和2.29%.实验结果表明,与国内外同类方法相比,采用多尺度注意力网络的立体匹配方法在匹配精度上获得了较大的提升,尤其是在弱纹理区域及物体边缘处表现较好.

Aiming at the problem that the depth of network is related to the size of training image patches and improving the matching accuracy for the weak texture and edge regions,a multiscale attention network for stereo matching is presented.The method is divided into two stages:in the first stage,a deep network for computing matching cost is proposed,which is composed of basic layer and scale layer,in the second stage,a disparity refinement network based on multiscale attention is proposed,in which multiple disparity clues are combined and multiscale attention is added to boost stereo matching accuracy.The percentage of 3-pixel bad points on KITTI 2012,KITTI 2015,and SceneFlow is 1.13%,1.87%,and 2.29%,respectively.Experiments show that compared with the current domestic and foreign advanced methods,a stereo matching method based on multiscale attention network made a great improvement in matching accuracy,especially better improvement for the weak texture and edge regions.