Monocular Depth Information Estimation And Super Resolution Recovery

Scene depth information is widely used in areas of computer vision,such as scene 3D reconstruction,human-machine interaction,autonomous driving,etc.However,the acquisition of depth information has always restricted the use of depth information.Current methods based on stereo vision can generate depth information,but it is often difficult to match the corresponding pixel and generate dense depth map.Depth map acquired from depth cameras also suffers from low resolution(LR)and noise,both are difficult to be applied in mobile devices.How to obtain high-quality and high-resolution depth map is currently a difficult problem to be solved.This paper first studied the problem of monocular depth information estimation.Considering objects have different scale size,this paper designed deep convolutional neural network(CNN)based on residual network to estimate the depth information between objects and camera from scene image.Besides,this paper used spatial pyramid structure which consists of multi-scale dilated convolutions to acquire object features of different scales and decrease the computational burden.And this paper proposed to use a depth enhancement subnetwork to enhance the low-resolution depth map due to the feature aggregation,to acquire higher quality depth estimation results.Experiments on indoor NYUv2 dataset and outdoor Make3 D dataset demonstrated the good performance of the proposed monocular scene depth estimation network.Then,as the depth map from depth cameras cannot meet the need of reality use,this paper studied the depth super resolution recovery to improve the quality of depth map.Considering the importance of edge information during the depth super resolution recovery procedure,this paper designed an edge-guided depth super resolution recovery framework,in which we first inferred the high-resolution(HR)edge map from LR depth map and corresponding color image,and then two methods were proposed for depth map super resolution recovery task,edge-guided local depth filling algorithm and an end-to-end super resolution recovery network.Middlebury dataset and MPI Sintel dataset are used to demonstrate the efficiency and generalization ability of two methods,it shows the proposed method can well preserve the sharpness of depth boundaries and increase the resolution and quality of depth map.It also shows that the proposed end-to-end super resolution recovery network perform well for noisy depth map super resolution recovery.