Research On Depth Estimation Based On Gaussian Depth Completion And Attention Mechanism

Recently,with the development of computer science and computational power,self-driving cars and relevant technique fetch much attention from industry and academia.Depth estimation plays a crucial role in 3D scene reconstruction which benefit car autonomous navigation and obstacle avoidance.Traditional Methods consume costly computational power and money,in the contrast,depth estimation methods based on deep learning represent huge potentiality in the future.The goal of depth estimation is to perform pixel-level regression and generate depth maps.The drawbacks depend on the pixel-level scale uncertainty of depth prediction and lack of semantic information which lead to indistinct depth edges and failure of small objects prediction.In order to solve those problems,we proposed a multi-stage learning strategy based on depth completion and attention mechanism for indoor and outdoor scenes,and we conduct research as follows:(1)This thesis proposes Self Adaptive Depth-Domain Gaussian Completion Module(SAGM).Under the premise of maintaining depth ordinal distribution,SAGM can augment the details of object edges and small objects through 3D spatial gaussian interpolation so that it generates more dense and accurate point cloud.It also alleviates the ill-posed restrictive problems of sparse distribution and foreground/background depth mix of datasets.(2)We designed Edge Aware Attention Module(EAAM)and Cascade Patch Mixed Attention Module(CPAM).The former has better semantic encoded ability which produces clearer depth edges and rich details of small objects.The latter aims at the ability of mixing multi scale cascade contextual patch of depth features.It also alleviates the size restraint of local reception fields of convolution kernel.(3)We proposed the whole framework named Multi Scale Cascade Residual Depth Net(MSCRD).We evaluate our model on three authoritative datasets such as KITTI,Make3 D and NYU Depth V2.The results show that MSCRD presents strong competitive behavior with respect to other state-of-the-arts approaches quantitatively and qualitatively.Meanwhile,the results also demonstrate our advantages of efficiency and generalization.