Efficient And Accurate Depth Estimation From Images

Depth perception plays an important role in 3D scene understanding.The rapid development of depth sensors makes it possible to obtain depth information directly through equipment collection,but they still have their own shortcomings.Kinect cam-eras are often only suitable for indoor environments,while LiDAR used in outdoor autonomous driving environments is quite expensive,and the depth data collected is relatively sparse,which limits their scope of use.Recovering the depth information of the scene from the images taken by the RGB camera presents a more versatile and inexpensive alternative.However,recovering depth from images is extremely challenging.Traditional methods generally need to solve a complex optimization problem,and at the same time,they may suffer in challenging situations such as textureless regions.In this paper,with the help of the efficient inference speed and the powerful representation learning ability of deep learning,the proposed method can efficiently obtain high-quality results.This paper first considers depth estimation from monocular videos using a deep learning approach.Existing methods can be broadly classified into supervised learn-ing and unsupervised learning.Supervised methods are often limited by small amount of ground truth depth dataset and unsupervised methods are mostly based on the static scene assumption,not performing well on real world scenarios with the presence of dynamic objects.To this end,this paper proposes a new deep learning method to es-timate depth from unconstrained monocular videos without ground truth supervision.Specifically,a bicubic function based deformation representation is proposed to model individual object motion.This representation enables our method to be applicable to diverse unconstrained monocular videos.We demonstrate the effectiveness of our pro-posed method on three public datasets.To obtain more robust depth estimation results,this paper further considers binoc-ular depth estimation.Current state-of-the-art stereo models are mostly based on costly 3D convolutions,the cubic computational complexity and high memory consumption make it quite expensive to deploy in real-world applications.We aim at completely replacing the commonly used 3D convolutions to achieve fast inference speed.To this end,we first propose a sparse points based intra-scale cost aggregation method to alleviate the well-known edge-fattening issue at disparity discontinuities.Further,we approximate traditional cross-scale cost aggregation algorithm with neural network layers to handle large textureless regions.With these two modules,we can not only significantly speed up(more than 40×)existing mothods,but also maintain competitive performance.Extensive results demonstrate the high efficiency and versatility of the proposed method.