结合注意力与无监督深度学习的单目深度估计

针对当前的无监督单目深度估计方法边界模糊的问题, 提出了一种基于双重注意力模块的网络架构。这种架构能有效利用图像特征的远程上下文信息解决深度估计中的边界模糊问题。整个框架使用基于视图合成的无监督方法训练, 模型框架包括深度估计网络与位姿估计网络, 同步估计深度和相机位姿变换。双重注意力模块嵌入在深度估计网络中, 包含位置注意力模块和通道注意力模块, 能表示远程空间位置和不同特征图间的上下文信息, 从而使网络估计出细节更好的深度信息。在KITTI数据集以及Make3D数据集上的实验结果表明, 本文的方法能有效提高单目深度估计的精度和解决深度估计边界模糊问题。

A Monocular Depth Estimation Combined with Attention and Unsupervised Deep Learning

To solve the problem of boundary blurring of current unsupervised monocular depth estimation method, a network architecture is proposed based on dual attention module. This architecture can effectively solve the problem of boundary blurring of depth estimation by using long-range context information of image features. The model framework that includes depth estimation network and pose estimation network is trained by an unsupervised method based on view synthesis and estimation depth and camera pose transformation at the same time. The dual attention module is embedded in the depth estimation network, including position attention module and channel attention module. This module can represent the long-range spatial location and the context information between different feature maps, so that the network can estimate the depth information with better details. The experimental results on the KITTI dataset and the Make3D dataset show that our method can effectively improve the accuracy of the monocular depth estimation and can solve the depth estimation boundary blur problem.