Sketch-based 3D Model Identification and Angle Estimation Using Cascaded LSDA

• Main Authors: Po-chung Lin, 林柏均
• Other Authors: Jenn-Jier Lien
• Format: Others
• Language: en_US
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/61625348224651215068

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 97 === Existing tools for reconstructing 3D models from user freehand sketch is a tedious work, as it requires manually choosing a targeted 3D model in a database and carefully matching 3D model to the desired 2D sketch. The major difficulty for automation of 3D reconstruction is that 3D model has various 2D contours caused by changing viewpoints. In this paper, we proposed a novel cascaded framework of 3D models reorganization and categorization for automatically choosing and matching tasks. In the training process, each 3D model in the database is decomposed as several 2D projected contours from different viewpoints. All contours are then organized in a cascade way combined with Locality Sensitive Discriminant Analysis (LSDA) to boost search efficiency. Also, manifold spaces are constructed to generate virtual 2D contours and consequently only a limited size of 2D contours is required in the database. In the testing process, the input free-form sketch is used for querying 2D projected contours from 3D database. The search stage is cascaded and parallel; at each layer, k-nearest neighbors of input sketch are selected and ranked by their similarity degree. The informative neighbors (only the top few of sorted list) are then used for indicating search direction in the next layer. Consequently, no user effort for choosing and matching 3D model is necessary where the object type and viewpoint are highly robust and efficiently estimated. Extensive experiments demonstrate that the proposed method is efficient and well-performed by testing for 8 object types, each has 1440 varied poses and 5 different contour ratios.