Depth Recovery For Objects In Single Images Using Exemplar 3D Models

Recovering 3D information from videos and pictures is a fundamental problem in computer vision,and has wide applications.For example,high-quality depth information can be used not only to reconstruct the 3D structure of a scene,and to assist in robot,navigation,but also to help the computer effectively identify and understand the objects and scenes in the images.Especially with the rise of appli-cations such as robots,virtual reality and augmented reality,it is becoming more and more important to make the computers better understand the 3D world based on the images or videos.In many cases,we can only have single images,so how to recover the depth map of a single image has important theoretical significance and practical value.With the rapid growth of 3D model data sets,a large quantity of researchers began to attempt to use 3D models for object depth recovery of single images.However,many existing methods require a high degree of matching between the input image and the data in the training set,which makes it difficult to deal with the difference between the objects in the input picture and the 3D models in the model library.Moreover,many methods need the result of segmentation as the priori knowledge for object depth recovery,which greatly restricts the practical applications.To solve these two problems,this paper develops an automatic ob-ject depth recovery system for single monocular images,which does not need extra input other than the query image.Firstly,the system selects images from training set to match with the query image to obtain the candidates,and adaptively divides these candidates into patches according to candidates' structural features.The use of patch matching result for depth recovery can avoid the disadvantages from matches between the whole image and training set.Secondly,the high-quality object segmentation and depth estimation are implemented based on the result of patch matching.Then,according to the condition that depth information ob-tained by matching is rough in the discontinuous boundary region of the object,the system utilizes the improved segmentation result to constrain the boundary and to fill the depth information of the missing pixels.Finally,the bilateral fil-tering is used to smooth the depth map to further improve the depth quality.A series of experimental results and comparisons with other methods demonstrate the effectiveness of the proposed method.