A 3D Model Retrieval Method Based On The Object Outline Of Image

The main content of this paper is a 3D model retrieval method based on the object outline of image. There are two classes of methods in this field, the methods based on key word and the methods based on content. Our method is one of the latter class. We use 2D image as a seed of retrieval. We extract the outline from object in image and the projection of 3D models. Then we compare these two kinds of character feature of outlines to retrieve 3D models.We built an entire 3D model retrieval system and all the work in this paper is a part of this project. The system is based on Web. It is running on Tomcat and the algorithms are implemented by Java. The IDE is JBuilder2006, the OS is Windows XP, and the CPU is P4 2.8G. Like other 3D model retrieval systems, our system presents three ways of retrieval. These are method based on key word, method based on 3D sample, method based on image sample. Users can choose prefer way by the interface. Besides, a user can feedback estimating information to system so the system learn that which 3D models of the result are satisfied by user while the others are not. The system will record all these information for advance process.In this paper, we use 2D outline distance method to compute characters of outline of object in image and the projection of 3D models. In this way we can simply describe the shape feature of outline. We compute distances from centre of the smallest enclosing circle to all N shape points on the outline. After this we get a vector which has N elements. Standardize result of this series of distances we can finally get the shape feature of the outline. We set the 5 points who have the minimum distance as the match-starting points. Therefore, since each character has 5 match-starting points, we have to compute 25 times in each comparison.The method we mentioned above is not completely rotation insensitive. And the complicated computation lowers its'efficiency. To resolve these problems, we use 2D Fourier Transform to extract feature from outline. First we find smallest enclosing circle of outline. Next we use N circles with the same centre of the smallest enclosing circle and M radials from the centre to divide the smallest enclosing circle into M(N+1) parts. Then we check each part weather it contains the outline or not and use 1 and 0 to denote. So we get a series of 1, 0 as input of 2D Fourier Transform. The absolute value of output is the feature we finally want.There is not such kind of 3D model feature which can be applicable in all cases at the present time. In another word, the ability of describing shape of different feature is different. In this paper we use feature combining technology to integrate two finds of features we mentioned above. We weight and combine two heterogeneous feature vectors. We have to compare each part of combined feature respectively, because the ways of comparison of different features are different. At last 16 related models could be returned.The experiments perform on the Princeton Shape Benchmark show that the proposed method can efficiently return the related models according to the images from various sources. The research not only is valuable for 3D model retrieval but also establish a new connection between the image database and the 3D model database. Further researches will perform on extracting object outline from image with complicated background and on improving the performance of outline's shape feature.