摘要
信息丰富且计算高效的代价体对于高精度高效率的立体匹配至关重要,为构建信息丰富且计算高效的代价体和实现高精度高效率的立体匹配,在Fast-ACVNet的基础上,提出一种轻量化网络Efficient-ACVNet,提升立体匹配在代价体构建阶段的效率。首先,使用计算量更低的3D代价体作为代价体注意力,引入逆瓶颈残差块堆叠对称沙漏结构对3D代价体进行代价聚合,并提出多尺度视差通道注意力模块来进一步提升代价聚合效果,聚合后的3D代价体作为代价体注意力将用于构建和滤波信息冗余的4D代价体,提升4D代价体的信息量和计算效率,最后引入伪3D残差块和设计伪3D下采样模块对4D代价体代价聚合,进一步降低网络复杂度。实验结果表明,与基准方法相比,所提算法在SceneFlow数据集上的端点误差(EPE)降低了9.375%,在KITTI15数据集上的前景区域(D1-fg)异常值百分比下降19%,并且网络运行时间从39ms下降到25ms。
An information-rich and computationally efficient cost volume is crucial for high-precision and high-efficiency stereo matching.To construct such a cost volume and achieve accurate stereo matching,a lightweight network,Efficient-ACVNet,was proposed based on Fast-ACVNet to improve the efficiency of cost volume construction in stereo matching.First,a computationally less intensive 3D cost volume was used as cost volume attention.Inverse bottleneck residual blocks were employed to stack a symmetric hourglass structure for cost aggregation of the 3D cost volume,and multi-scale disparity channel attention modules was introduced to further enhance the aggregation effect.The aggregated 3D cost volume served as cost volume attention to construct and filter the information-redundant 4D cost volume,improving its information content and computational efficiency.Finally,pseudo-3D residual blocks were introduced and pseudo-3D downsampling modules was designed for cost aggregation of the 4D cost volume,further reducing network complexity.Experimental results showed that compared to the baseline method,the proposed algorithm reduced the endpoint error(EPE)by 9.375%on the SceneFlow dataset,decreased the outlier percentage(D1-fg)in the foreground region by 19%on the KITTI15 dataset,and reduced network runtime from 39 ms to 25 ms.
作者
李冰
严熠萌
张鑫磊
邵宝文
翟永杰
Li Bing;Yan Yimeng;Zhang Xinlei;Shao Baowen;Zhai Yongjie(Department of Automation,North China Electric Power University,Baoding 071003,China;Baoding Key Laboratory of Intelligent Robot Perception and Control for Power Systems,Baoding 071003,China)
出处
《电子测量技术》
北大核心
2025年第20期179-188,共10页
Electronic Measurement Technology
基金
国家自然科学基金面上项目(62373151)
国家自然科学基金联合项目(U21A20486)
河北省自然科学基金面上项目(F2023502010)
中央高校基本科研业务费专项资金(20237488)资助。
关键词
立体匹配
深度学习
代价体注意力
逆瓶颈残差块
伪3D卷积
stereo matching
deep learning
cost volume attention
inverted bottleneck residual block
pseudo-3D convolution
