一种基于空间上下文注意力网络的立体匹配方法

Research on the Stereo Matching Method based on Spatial Context Attention Network

立体匹配可以看作是有监督的学习任务,通过将大量的左右图像输入到卷积神经网络中进行训练,可以获得性能良好的视差图。但是,当前的网络结构对病态区域的视差估计仍然存在困难。为了解决这个问题,提出了一种空间上下文注意力网络,这个网络可以通过聚合全局上下文信息来提高对病态区域的视差估计准确性。实验结果表明,提出的方法在合成数据集Scene Flow和真实场景数据集KITTI 2015中对病态区域的视差估计得了较好的效果。

Many works have shown that stereo matching can be regarded as a supervised learning task.By inputtinging a large number of left and right images into the convolutional neural network for training,a good disparity map can be obtained.However,the current network structure still has difficulties in the disparity estimation of the ill-posed regions.To solve this problem,a spatial context attention network was proposed that can improve the accuracy of disparity estimation in ill-posed regions by aggregating global context information.The experimental results show that the proposed method achieves good results in the synthetic dataset Scene Flow and the real scene dataset KITTI 2015.