Unsupervised Monocular Depth Estimation

Estimating the depth information of monocular images is a classic and challenging problem.In traditional computer vision,sparse depth values can only be calculated from binocular images.Due to the scale uncertainty of monocular images,depth values cannot be calculated only from monocular images.At present,the monocular depth estimation method based on deep learning technology has also become a research hotspot.Monocular depth estimation methods based on supervised learning usually rely on a large amount of true depth for model training,and the cost of obtaining the true depth is very high.The monocular depth estimation based on unsupervised learning is usually trained according to image reconstruction loss and does not need to rely on the depth truth value.To solve the problem that the monocular depth estimation based on supervised learning rely on the real depth,the monocular depth estimation based on unsupervised learning is proposed in this paper,and the main work is as follows:Firstly,an unsupervised depth estimation method based on SR-BINet encoder-decoder network is proposed to solve the problem of invalid feature influence model optimization in monocular images and the checkerboard effect caused by deconvolution.First of all,the encoder of SR-Binet network adds learnable weight to each layer of features,increases the weight of effective feature mapping in network training,and reduces the weight of ineffective or ineffective feature mapping.Therefore,the encoder of SR-Binet network can extract effective features more efficiently,and remove the interference of invalid features in the image,which makes the network model more easily converges.Secondly,the decoder of SR-Binet network samples the low-resolution features,and reverts the low-resolution feature maps to high-resolution parallax values by bilinear interpolation and convolution,which can eliminate the influence of the checkerboard effect brought by deconvolution.Secondly,to solve the occlusion problem in image reconstruction,an unsupervised depth estimation method based on adaptive left-right consistency constraint is proposed to reduce the influence of occlusion region on model optimization.During image reconstruction,the image reconstruction error of the occluded area and the non-shared area is relatively large,while the reconstruction error of the shared area is relatively small.When calculating the left and right consistency loss,the two areas need to be considered separately.Therefore,this paper proposes adaptive left-right consistency constraints and error distributions for balanced occluded and unoccluded regions.The global average error of image reconstruction is calculated by using L2 norm,and a weight calculation method of single pixel error relative to global average error is designed.The weight calculation is related to the image reconstruction error during training,and the weight can be calculated adaptively according to the size of the error,so as to achieve the purpose of balancing the error distribution.Finally,the experiments are carried out on the KITTI dataset.The experimental environment,parameter settings and experimental results are described in detail in this paper.Compared with other monocular depth estimation methods,the experimental results of this paper are the state-of-the-art,so it proves the effectiveness of the method proposed in this paper.