基于多尺度残差掩码的双目立体匹配算法

Stereo Matching Network Based on Multiple Scales Residual Refinement Mask

基于卷积神经网络结构的立体匹配算法已有较好表现,但仍存在以下局限:基于卷积结构的特征提取器利用的先验信息较少,特征相关联性不足,致使病态区域对算法产生较大干扰;静态的卷积核难以平衡差异化的特征,带来过度平滑的效果。文章提出的多尺度残差特征掩码模块从两方面可以改善以上问题:利用左图先验信息引入丰富的语义帮助调优算法选取相关度高的邻居点;利用右图与调优前的视差重建左图并与左图对比获取残差,自适应地依据残差区域尺度选择合适的邻居点个数,摆脱误差区域干扰。将文章提出的方法集成到PSM-Net基本框架中,在KITTI2015 Stereo数据集上的实验结果表明,所提出的方法有效提升了算法在病态区域上的表现。

The stereo matching algorithm based on the convolutional neural network structure has performed better,but there are still the following limitations:the feature extractor based on the convolutional structure utilizes less a priori information about the feature phase correlation,which makes the pathological regions interfere more with the algorithm;the static convolutional kernel is difficult to balance the differentiated features,which brings the effect of over-smoothing.The multiscale residual feature mask module proposed in the article improves the above problems from two aspects:using the a priori information of the left map to introduce rich semantics to help the tuning algorithm select neighbors with high correlation;using the parallax between the right map and the pre-tuning view to reconstruct the left map and compare it with the left map to obtain the residuals,adaptively selecting the appropriate number of neighbors based on the scale of the residual region to get rid of the interference of the error region.The experimental results on the KITTI2015 Stereo dataset by integrating the proposed method into the basic PSM-Net framework in the paper show that the proposed method effectively improves the performance of the algorithm on the pathological regions.