Research On 3D Model Retrieval Based On Supervised Bag-of-Visual-Words Framework

3D models, as a fourth type of multimedia data besides sounds, images and video, have been widely used in many fields, such as 3D television animation, mechanical design, geographic information system, etc. Computer software and hardware’s great improvement substantially increases the types and number of 3D models. In order to promote the work efficiency and reduce the development costs, a large number of 3D model databases appeared in various fields to realize the reuse of resources. Moreover, the popularity of the Internet has further accelerated the dissemination and sharing of data. Therefore, it is of great importance to study how to retrieve the required data accurately from a large number of 3D models, which has strongly significant theoretical and practical value.In the light of 3D models’ shape attribute, this paper does some researches on visual-based 3D retrieval approaches, and proposes a novel 3D model retrieval algorithm based upon supervised bag-of-visual-words. It mainly includes the following three parts. Firstly, when projecting a 3D model in multi-view, an adaptive visual distance rendering operation is proposed, so that the rendered depth-buffer images can retain the model’s spatial shape information to the maximum extent. It adaptively adjusts the projection plane to the tangent plane of the 3D model in a certain view angle, resulting in rendered depth images with more abundant texture information, which are easier to distinguish. Secondly, in order to make full use of the significant category information of the classified database, a novel 3D model feature encoding method is presented based on the supervised bag-of-visual-words framework. The traditional universal vocabulary is replaced by the supervised class vocabularies, getting some more complete vocabularies to depict the local typical characteristics of a 3D model, which makes the follow-up feature vector more accurate to describe the shape of a 3D model. Finally, as we all know, in practical application, users are more concerned about the search results in front pages that those in latter. Motivated by this, the thesis puts forward a retrieval results correction strategy. The category labels of the initial search results’ first few pages are statistically analyzed, and accordingly, the classification of the input model is determined. And then the initial query results are corrected based upon the former analyzed verdict, which further promotes the final retrieval accuracy.Experiments are operated on the Princeton Shape Benchmark, and the results show that the proposed adaptive visual distance rendering and supervised class vocabularies based 3D model retrieval algorithm upgrades the mean average precision to 74.1%, which is higher than the existing unit sphere rendering and universal vocabulary based 3D retrieval approach by 18.1%. It demonstrates that the presented method can describe and distinguish the shape of 3D models more accurately and effectively. In addition, after revise the initial query results employing the raised correction strategy, the retrieval precision is greatly enhanced to 90.1%, validating the effectiveness of the correction operation. Besides, because of the independence between the revision operation and the prior retrieval process, the retrieval results correction strategy can be directly popularized to other 3D retrieval approaches, which can further boost their final retrieval accuracy.