Monocular Image Depth Estimation Based On Transfer Learning

How to obtain three-dimensional spatial information of image from two-dimensional image is one of the important research contents in computer vision.The depth estimation of the image becomes the research hotspot of computer vision,because the depth information is an important clue for the expansion from the 2D planar structure to the three dimensional information.In this paper we presents a depth estimation algorithm based on migration learning.Firstly,the convolution neural network is used to extract the features of the input image and the database image.Secondly,the candidate image is drawn according to the feature similarity of the input image and the database images by cosine similarity.Thirdly,the depth information of the candidate image is deformed again by the SIFT stream so that the input image is aligned with the candidate image.Fourthly,construct the objective function of optimizing the depth of deformation candidate.Finally,the IRLS algorithm is used to optimize the objective function iteration till it converges,and the final depth map of the input image is output.The experimental results show that the depth estimation of the single image can be realized by using the convolution neural network for transfer learning.In this thesis,we extracted the feature by the convolution neural network and estimated depth information by transfer learning.The proposed method can effectively estimate the depth of a single image and video sequence.