基于多尺度特征及Wasserstein距离损失的单目图像深度估计

Depth Estimation of Monocular Images Based on Multi-scale Features and Wasserstein Distance Loss

单目图像深度估计是一个病态问题,究其原因在于单一图像中缺乏深度信息。随着深度学习技术的不断发展,深度神经网络在单目图像深度估计领域取得了一定的突破。现有的深度网络模型采用编码-解码结构,编码器往往采用全卷积的方式来获取特征图像,但提取的特征图像往往不能很好地反映图像原有的信息。因此,通过对图像编码器进行改进,对提取的不同尺度下的特征进行融合,使得特征图像能更好地反映原有的图像信息。并且在训练时,在左右一致性损失的基础上引入Wasserstein距离损失对模型进行约束。实验结果表明,模型在KITTI数据集上具有较好的表现,预测得到的深度图像具有较高的准确性。

Monocular image depth estimation is a pathological problem,the reason lies in the lack of depth information in a single image.With the development of deep learning technology,deep neural networks have made a breakthrough in depth estimation of monocular images.Existing deep network models adopt coding-decoding structure,and encoders often adopt full convolution to obtain feature images,but the extracted feature images often cannot reflect the original information of images well.Therefore,by improving the image encoder,this paper fuses the extracted features at different scales,so that the feature image can better reflect the original image information.In addition,Wasserstein distance loss is introduced to constrain the model on the basis of left and right consistency loss.The experimental results show that the model performs well on the KITTI dataset,and the predicted depth image has high accuracy.