一种结构感知损失的域适应深度估计方法

A Domain Adaptive Depth Estimation Method for Structural Perception Loss

针对域适应深度估计方法中域图像间结构差异较大问题,文中提出一种结构感知损失的域适应深度估计方法。该方法通过预训练的卷积神经网络对图像进行特征提取,在特征上进行结构相似性度量,减小了域图像之间的差异性,提高了转换模块的稳定性。该方法使用合成图像深度对和真实图像训练,不需要真实图像的深度标签和物理几何信息。在KITTI数据集上进行实验,深度准确率达到了96.6%,证明该方法可有效提高深度准确率。

In the domain adaptive depth estimation method,the structural differences between domain images are large.In view of this problem,a domain adaptive depth estimation method for structural perceptual loss is proposed.This method uses pre-trained convolutional neural networks to extract features from images,and measures structural similarity on features,which reduces the difference between domain images and improved the stability of the transform module.This method uses synthetic image depth pairs and real image training,and eliminates the requirement of depth labels and physical geometric information for real images.Experiments on the KITTI dataset achieve a depth accuracy rate of 96.6%,which proves that the method can effectively improve the depth accuracy.