实时自适应的立体匹配网络算法

Real-time Adaptation for Stereo Matching

立体匹配是一个经典的计算机视觉问题。采用传统方法或卷积神经网络(CNN)方法的立体匹配,其精确度和实时性不能满足实际的在线应用。针对该问题,本文提出一种实时自适应的立体匹配网络算法,通过引入一种新的轻量级的、有效的结构模块自适应立体匹配网络(Modularly Adaptive Stereo Network,MASNet),在网络中嵌入无监督损失模块和残差细化模块,使立体匹配的准确性和实时性得到提高。实验结果表明,本文方法相比具有相似复杂度的模型,精确度更高,并且能以平均约25帧每秒的处理速度达到在线使用的要求。

Stereo matching is a classical computer vision problem.The accuracy and real-time performance of the traditional method or convolutional neural network(CNN)method for stereo matching cannot meet the requirements of online application.Therefore,a real-time adaptive stereo matching network algorithm was proposed in this paper.The accuracy and real-time performance was improved by introducing a new lightweight and effective architecture Modularly Adaptive Stereo Network(MASNet),embedded an unsupervised loss function and residual refinement module.Firstly,the multi-scale features are extracted by the pyramid network.Then the initial disparity estimation is carried out,and finally,the final disparity map is output by the residual refinement module.The experimental results show that the proposed method is more accurate than the model with similar complexity,and the processing speed of about 25 frames per second on average meets the requirements of online usage.