Research On 3D Model Acquisition And Auto-tagging Algorithm

ABSTRACT:In With the rapid development of three-dimensional technology,3d models are widely applied in virtual reality, architectural design,3D animation, biomedical modeling and other areas. However, the modeling process of 3d model is laborious and time-consuming, and the 3d model files available are lack of semantic annotation information for understanding. So, more researches on the semantic annotation technology of 3d models are urgently needed, in order to easily search 3d models and reuse them. Therefore, aimed at the problem that the 3d model available on the network currently are lack of semantic annotation information, and that tagging the models by manual work is laborious and heavy workload, as well as that the 3d model libraries for retrieval algorithm research contain only small number of models and need to be expanded, we do lot of work on the research of 3d model retrieval methods, the expansion of 3d model training database, and the research of 3d model auto-tagging algorithm in this paper.The main research contents of this paper are as follows:1. Through the comparison of several 3d model retrieval methods, we find that the text retrieval has low search accuracy, and the shape retrieval has poor user's interaction. So, we propose a 3d-shape retrieval method to improve the text retrieval. Firstly, do the text retrieval based on Google 3D Warehouse, then download the retrieved models and convert SKP model to OFF model, finally re-sort the search results according to the models'DESIRE feature vectors. The DESIRE feature vector is a composite 3D-shape feature vector, which is formed using depth buffer images, silhouettes, and ray-extents of a polygonal mesh. According to certain weight related to their retrieval performance, these three features are combined up to a much better description of 3d model shape. In this method, text retrieval and shape retrieval are combined to improve the text search accuracy, and have a good user's interaction.2. The commonly used 3d model training set, such as PSB model set provided by Princeton University, is too small, and owns only 1814 models, and the model number of each semantic category is unequal. However, the 3d model library on line, such as Google 3D Warehouse, has a great number of models published by users for free, but lacks classification information. More, the models in them are not well labeled and the model file format is not suitable for retrieval algorithm research. Therefore, by means of the model retrieval system mentioned in the first part, we take the label name of the semantic category which has fewer models as the keyword and do the text retrieval. Then, download the retrieved models from Google 3D Warehouse and covert their file format to OFF. Finally, expand the PSB model set with the collected models, and establish a more comprehensive 3d model training database.3. According to the analysis of the existing 3d model's tagging algorithms, we propose a 3d model semi-automatic annotation method based on fuzzy K nearest neighbors, in order to solve the problem of 3d model's lack of semantic information and tagging the models by manual work is laborious and heavy workload. This method can automatically tag the models. Our method uses a fuzzy KNN classification algorithm, and classification results are fuzzy, and the results contain more models'information. Then, by calculating the uncertainty of the fuzzy classification results, we introduce an active learning mechanism. This mechanism can semi-automatically tag the unlabelled model set through the steps of threshold comparison, user's relevance feedback and multiple iteration learning. Experiments show that the method can complete the work of annotation well and get better classification accuracy and further improve the automation degree of tagging the model.