SKETCH-BASED 3D MODEL RETRIEVAL FOR GAME DEVELOPMENT DOMAIN

• Main Author: Aji Nugroho, Adam
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39142
• Institution: Institut Teknologi Bandung

Game heavily used multimedia data such as 3D models for graphic purposes. The increasing number of 3D models developed, the more difficult it is to organize files. As a result, it is more difficult to search for existing 3D models using metadata file such as file name. Similarity-based 3D model retrieval system can help to search for 3D models without having to use the corresponding file name. However, the 3D models retrieval system that have been developed only considers mesh while 3D models in games have low-poly characteristics so that different 3D models can have similar meshes. This can lead to ambiguity in search. Looking at these problems, a 3D model retrieval is developed to handle low-poly in 3D game models. 2D digital sketches are used to retrieve similarities with 3D models. A 3D model will be sampled into 2D with the dodecahedron model and feature extraction will be done using several methods. There are 3 types of features used, global, local and texture. Global is based on shape features, local used gradient features from image patches and texture for extraction of texture data in 3D models based on surface patterns and colors, so that retrieval engine also consider textures not only mesh. The three features are combined into a composite feature to expand the characteristics of the features obtained. Global as a filter, while local and texture to refine the query. 3D model datasets are collected from various game development sites to obtain low-poly and textured characteristics. Digital sketch datasets were also developed to test retrieval engine. From the analysis in the testing process, it was found that the 3D model retrieval engine was influenced by sketch types, inter-class variation, intra-class similarity and combining features such as Distance Integration can improve performance. The test results from developed retrieval engine are quite well for the 3D game model with the first 10 ranks, the average precision is 0.827, nearest neighbor 1, 1st tier 0.848, 2nd tier 1.289 and discounted cumulative gain 0.973.