融合多层次特征及互注意力机制的视差估计  

Disparity estimation based on fusion of multi-level feature and mutual attention mechanism

作　　者：韩坤昊 贾振堂 Han Kunhao;Jia Zhentang(College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 201306,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海201306

出　　处：《国外电子测量技术》2023年第10期35-42,共8页Foreign Electronic Measurement Technology

基　　金：国家自然科学基金(62105196)项目资助。

摘　　要：提出一种端到端的基于互注意力机制并融合多级特征的视差估计算法,利用多级特征图像产生多个视差代价空间,以当前视差对这些代价空间进行检索,将检索结果融合后形成代价特征;同时利用左特征图构建内容特征。代价特征和内容特征在互注意力模块中多次相互作用后输出综合特征,并据此拟合视差。互注意力模块可多次迭代,以进一步优化视差。与RAFT-Stereo算法相比,算法的网络参量降低了23%,运算速度提高了76.2%,并在KITTI 2015数据集上的终点误差(EPE)降低了0.15,验证了该方法的有效性。This paper presents an end-to-end disparity estimation algorithm based on mutual attention mechanism and multi-level features is proposed.Multiple disparity cost spaces are generated by using multi-level feature images.These cost spaces are retrieved by the current disparity,and the retrieval results are fused to form cost features.At the same time,the left feature graph is used to construct the content feature.The cost feature and content feature interact with each other several times in the mutual attention module to output the comprehensive feature,and then fit the disparity accordingly.The mutual attention module can be iterated many times to further optimize the disparity.Compared with RAFT-Stereo algorithm,the network parameters of this algorithm are reduced by 23%,the operation speed is increased by 76.2%,and the EPE error on KITTI 2015 data set is reduced by 0.15,which verifies the effectiveness of this method.

关 键 词：视差估计 互注意力机制 余弦相似度 多级融合 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]