| Monocular depth estimation is an important research topic in computer vision,and it has broad application prospects in many fields such as autonomous driving,robot navigation,and3 D modeling.In recent years,monocular depth estimation based on unsupervised learning attracts continuous attention due to the elimination of expensive depth annotations,which uses stereo videos,monocular videos,or stereo pairs during training and restores the pixel-wise depth map from a single image during testing.Although unsupervised monocular depth estimation methods have made great progress,they still have some problems.First,existing unsupervised monocular depth estimation methods based on stereo videos do not make use of the advantages of stereo data,whose performance needs to be improved.Secondly,many unsupervised monocular depth estimation methods use complex network architecture or additional auxiliary information,which increase the cost and burden of learning.Finally,these unsupervised monocular depth estimation methods training from stereo pairs still have the problem of inaccurate photometric reconstruction loss.Aiming at the above-mentioned problems in unsupervised monocular depth estimation,this thesis mainly studies the following works:(1)An unsupervised monocular depth estimation framework based on two-stage and multitask learning is proposed.First,the two-stage multi-task unsupervised training strategy is used to make full use of the advantages of binocular depth estimation,thereby improving the performance of monocular depth estimation and ego-motion estimation.In addition,a dense feature fusion module is proposed to further enhance the characterization ability of the monocular depth estimation network without increasing the amount of network parameters.Qualitative and quantitative experiments on multiple public datasets demonstrate that the proposed method achieves superior performance on both monocular depth estimation and egomotion estimation compared with existing unsupervised methods.(2)An efficient unsupervised monocular depth estimation method training from monocular videos is proposed.This method explores the potential of unsupervised monocular depth estimation from three aspects: data augmentation,network architecture,and loss function.First,a simple and effective data augmentation method is proposed to increase the number and diversity of training samples,thereby enhancing the generalization performance of the model.Secondly,an efficient and low-overhead network architecture is proposed,which can improve the performance of the depth estimation model while reducing the amount of network parameters and network calculations.Finally,a novel scene depth consistency loss function is proposed to guide the model to dig out more useful depth cues,thereby improving the robustness of the model.Comprehensive experiments on multiple public datasets demonstrate the effectiveness of the proposed method,and the proposed model achieves the state-of-the-art results.(3)An unsupervised monocular depth estimation method based on multi-mask enhancement is designed.These unsupervised monocular depth estimation methods training from stereo pairs have a problem of inaccurate photometric reconstruction loss,which is mainly caused by the occlusion areas and uncertainty areas such as low texture or transparent surface between the stereo pairs.Therefore,the proposed method designs a binary occlusion mask to locate the occlusion areas between the stereo pairs,and meanwhile learns an uncertainty mask to deal with the uncertainty areas.Finally,these two masks participate in the image reconstruction process,alleviating the problem of inaccurate photometric reconstruction loss caused by occlusion areas and uncertainty areas.The experiments on the public datasets demonstrate the effectiveness of the proposed method. |
