摘要
立体匹配是从一对左右视图中估计视差(或深度)的过程,是计算机视觉中的基础问题。现有大多数立体匹配方法采用基于卷积神经网络(CNN)的深度学习技术,其中代价体构建与代价体成本聚合对于立体匹配的精度和效率至关重要,但现有方法需要在精度与速度之间作出平衡。针对该问题,加入注意力机制构建代价体,引入上下文特征—几何特征融合(CGF)方法进行代价聚合,借鉴Fast-ACVNet模型的基础结构提出一种新的视差分析网络CG⁃Fast-ACVNet。注意力代价体利用相关代价体中编码的相似性信息来正则化连接代价体,从而实现了整体的高效性和高准确性。该方法使得仅需轻量级的聚合网络即可达到理想效果,且CGF能够自适应地融合上下文和几何信息,以实现更有效的代价聚合。同时,为特征学习提供反馈,指导更高效的上下文特征提取。实验结果表明,该方法在Scene Flow、KITTI和ETH3D数据集上均取得了很好的效果。模型充分利用了两种代价体的优势,在显著降低代价体构建时间以获得实时性的同时,准确度也有显著提升,并具备较强的泛化能力。
Stereo matching,the process of estimating disparity(or depth)from a pair of left and right views,is a fundamental problem in com⁃puter vision.Most existing stereo matching methods use deep learning techniques based on convolutional neural networks(CNNs),where cost volume construction and cost aggregation are crucial for accuracy and efficiency.However,these methods often require a balance between pre⁃cision and speed.To address this issue,we incorporate an attention mechanism to construct the cost volume and introduce a Context-Geometry Feature fusion(CGF)method for cost aggregation.Building on the Fast-ACVNet model architecture,we propose a new disparity analysis net⁃work,CGFast-ACVNet.The attention concatenation volume leverages similarity information encoded in the correlation volume to regularize the concatenation volume,achieving overall high efficiency and accuracy.This approach allows a lightweight aggregation network to achieve optimal results.The CGF method adaptively fuses contextual and geometric information for more effective cost aggregation,providing feedback for feature learning and guiding more efficient contextual feature extraction.Our method achieves excellent results on the Scene Flow,KITTI,and ETH3D datasets.Experiments demonstrate that the model effectively utilizes the advantages of these approaches,significantly reducing cost volume construction time while achieving real-time performance and improving accuracy,with strong generalization capabilities.
作者
李瀚灵
黄影平
LI Hanling;HUANG Yingping(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《软件导刊》
2026年第2期36-45,共10页
Software Guide
基金
国家自然科学基金项目(62276167)
上海市自然科学基金项目(20ZR1437900)。
关键词
立体匹配
匹配代价体
代价体聚合
计算机视觉
stereo matching
matching cost volume
cost volume aggregation
computer vision
