Stereo Matching Based Multi-scale Context And Attention

Stereo matching is a basic and important research topic in computer vision.Its task is to calculate the disparity of each pixel in the reference image based on a rectified pair images.Recently,with the development of convolutional neural network,stereo matching based on deep learning achieves remarkable performance.However,they still fail to find accurate corresponding points in ill-posed regions such as repetitive texture,textureless and occlusion areas.In this paper,we propose a stereo matching algorithm based on multi-scale context and attention,based on DispNets and the characteristics of stereo matching.Firstly,we propose a stereo matching algorithm based on multi-scale context information,which uses spatial pyramid pooling layer to fuse multi-scale context information.And hole convolution maintains high spatial resolution in the downsampling stages,reducing the missing of high-level semantic information of small-scale objects.Secondly,we propose a multi-scale attention algorithm that enables the network to provide pixel-level attention information adaptively for disparity prediction and reducing redundant information based on residual attention.Finally,we propose a final stereo matching algorithm based on multi-scale context information and attention mechanism,which combines the pyramid pooling layer and attention.We test and compare with some state-of-the-art methods on the Flyingthings3 D,KITTI2012 and KITTI 2015.The proposed multi-scale context and attention mechanism-based stereo matching algorithm significantly outperformed the baseline DispNets in the three datasets.Finally,our stereo matching algorithm based on multi-scale context and attention achieves more accurate result,and our method is also competitive and effective compared with others in the same phase.