Content-Based 3D Model Retrieval

As the ever-increasing requirements of modeling speed and quality in multimedia application field, design and model reuse is the key to reducing product development time and cost. Since the traditional search systems employing keywords have had limited success rates, considerable emphasis has been placed over content-based 3D model search technique. The focus of our research in this paper is on the interactive query interface based on photos and model feature descriptors based on visual similarity, the main work can be conclude as following:(1) The distortion problem of visual similarity feature is analyzed and modifiedThe descriptors based on visual similarity achieve best performance comparing with other content-based 3D model retrieval techniques have been proposed in recent years. The descriptors based on visual similarity represent object by the collection of 2D shapes obtained from the projection of 3D model in different views. However, the methods based on 2D visual similarity just apply invariant shape features directly, which distorts model feature space and leads to incorrect matches. Taking Light Field Descriptor (LFD) proposed by National Taiwan University as example, we analyze the disadvantage of model descriptors based on visual similarity and illustrate its impact to retrieval accuracy. In this paper, we propose a kind of modified descriptor which preserves the exact geometrical features of the model in each view, therefore visual differences between models belong to different classes is hold. Our experiment results demonstrate that, the modified Light Field Descriptor achieve 5.3%, 4.1%, 3.6% and 2.3% better performance in term of First-Tier, Second-Tier, E-Measure and Discounted Cumulative Gain measures than the original approach.(2) A photo-based interactive query method is proposedContent-based model retrieval systems retrieve objects based on the shape similarity between objects, however it is difficult for user to get satisfactory queries through simple and convenient operation by existing query methods which would lead to poor results. We developed a new query interface to assists user to sketch precise 2D shape query from ordinary photo. The proposed user-friendly interface offers a time saving and accurate query method for users. Since the query interface shows the dynamic contour automatically in real time for reference, the contour and shape achieved from the proposed interface could keep the geometrical features which are identical to actual object. Using the new 2D sketch interface, the whole query process only takes several seconds which improves the effectiveness of retrieval process dramatically.