多尺度信息引导代价体金字塔聚合的立体匹配网络  

Stereo Matching Network Based on Multi-Scale Information Guided Cost Volume Pyramid Aggregation

作　　者：惠康华[1] 张榆 高思华 HUI Kanghua;ZHANG Yu;GAO Sihua(College of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China;College of Computer Science and Technology,Jilin University,Changchun 130012,China)

机构地区：[1]中国民航大学计算机科学与技术学院,天津300300 [2]吉林大学计算机科学与技术学院,吉林长春130012

出　　处：《昆明理工大学学报（自然科学版）》2024年第6期64-74,共11页Journal of Kunming University of Science and Technology(Natural Science)

基　　金：国家重点研发计划基金项目(2020YFB1600101);中央高校基本科研业务费中国民航大学基金项目(3122015C021).

摘　　要：针对基于4D代价体金字塔的立体匹配网络存在多尺度信息利用率低,从多代价体中提取关键信息困难等不足,提出一种多尺度信息引导代价体金字塔聚合的立体匹配网络.为有效利用代价体金字塔聚合过程中不同尺度的输出信息,设计一种基于单峰视差注意力的尺度间聚合方式,通过代价体金字塔上部分的预测视差突出金字塔底层代价体中的重点信息;为增强代价聚合对于显著特征的关注,提出引导沙漏结构,通过图像的多尺度几何信息与代价体的全局空间信息,自适应地聚合与校准各尺度代价体.实验结果表明,该网络在SceneFlow、KITTI 2012和KITTI 2015数据集上取得良好匹配准确度的同时具有较快推理速度,且与CFNet和PCWNet相比,参数量分别降低31%与57%.To solve the problem of insufficient utilization of multi-scale information in the existing stereo matching networks based on a 4D cost volume pyramid.A stereo matching network based on multi-scale information guided cost volume pyramid aggregation is proposed,this network utilizes multi-dimensional attention mechanisms to optimize the cost volume pyramid aggregation module to improve the utilization of multi-scale information.The network uses the initial prediction of the cost volume pyramid as the attention of the disparity dimension to filter the cost volume and designs a guided hourglass architecture that uses multi-scale feature maps and the global spatial information in cost volume to generate attention for cost volume on the channel dimension.The above methods improve the utilization of multi-scale information by enhancing the interaction between different scales of information,thus improving the network's ability to discern important information.The experimental results show that the proposed network achieves a good balance between inference speed and matching accuracy on SceneFlow,KITTI 2012 and KITTI 2015 datasets,and reduces the number of parameters by 31%and 57%compared to CFNet and PCWNet,respectively.

关 键 词：立体匹配 注意力机制 多尺度信息 代价体金字塔 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]