Research On Sketch Based 3D Shape Retrieval

In recent years,with the improvement of computer hardware and the rapid development of multimedia technology,more and more 3D models are used in file production,3D printing,game making and so on.Considering explosive growth of the amount of 3D models,the study of 3D model retrieval technology is of great significance.Traditional retrieval methods based on keyword and category cannot meet the demand for large-scale 3D model database because a lot of 3D models are difficult to describe in words and the amount of 3D models are too large to find the desired model in a specific category.Sketch-based query method can provide more information than text while it's easy to use.To solve the cross-domain problem for sketch based 3D shape retrieval,3D models are rendered into line drawings in different viewpoints.Because of the random placement of 3D models,we need to select the best views from different viewpoints to representative features for 3D models.Moreover,there are big differences between line drawings and freehand sketches.In this research,we deeply studied sketch-based 3D shape retrieval,and proposed an efficient retrieval framework.The main work and contributions are as follows:(1)Present a sketch-guided view selection method for 3D models.First,a sketch classifier based on convolutional neural networks is built and trained.Then a probability-based view selection method is proposed based on this classifier.Finally rank lists for views of each 3D model are generated and best views are obtained for 3D model feature representation.(2)Present a deep metric learning network for sketch-based 3D shape retrieval.First,the feature extraction module,which are part of the classifier,is used to extract features for both sketch and views of 3D models.Then a deep metric learning network is built to map features from different domain into the same domain by learning the correlations between sketch and views of 3D models.Therefore,the performance of cross-domain is improved.