基于两阶段代价矩阵和动态注意力的双目立体匹配网络

EDNet++: Improving Stereo Matching with Two-Stage Combined Cost Volume and Multiscale Dynamic Attention

摘要:

目前大多数先进的双目立体匹配网络通过构建4D代价矩阵以保留图像的语义信息，增加了网络的计算量开销。为了解决上述问题，提出了两阶段的组合代价矩阵和多尺度动态注意力的EDNet++网络。首先从全局的、粗粒度的视差搜索范围上构建的基于相似度的代价矩阵作为引导，在局部的搜索范围上实现细粒度的组合代价矩阵，其次提出基于残差的动态注意力机制，其根据中间结果信息自适应地生成空间上的注意力分布，并且通过迁移实验证明了该方法的有效性，最后在各大公开数据集上的对比实验结果表明，相较于其他方法，EDNet++方法能够达到算法精度和实时性的良好平衡。

Abstract:

Most state-of-the-art stereo matching networks construct 4D cost volume to preserve the semantic information of the image， which increases the computational cost of the network. To solve this problem， a network named EDNet++ with a two-stage combined cost volume and a multi-scale dynamic attention is proposed. First， a correlation cost volume is constructed based on global and coarse-grained disparity search range， which is used as a guide to construct a fine-grained combined cost volume on the local disparity search range. Then， the dynamic attention mechanism based on residuals can adaptively generate spatial attention distribution according to the intermediate result information， and the effectiveness of this method is proved by the transfer experiment. The comparison experiments on various public data sets show that EDNet++ can achieve a good balance between accuracy and real-time performance compared with other methods.