3D Modeling Based On Shape Understanding

3D modeling is one of the most important problems in Computer Graphics.In the information era,the demand for 3D digital models is growing.Nowadays,most of 3D models are created by professional designers and modelers.it's not easy for novice users to acquire high-quality 3D data,or adjust the details of the 3D models to ensure the quality.Therefore,the main goal of 3D modeling is to assist non-professional users to create high-quality 3D models easily and efficiently.Common 3D models include objects and scenes,whose quality is affected by the geometric properties of local surfaces(continuity,smoothness,compactness,etc.)and the plausibility of the global shapes(whether the object shapes meet the semantic mean-ing,or whether the object relations support the functional and interactive goals).The understanding on different aspects undoubtedly can help to improve the quality of the 3D models.Especially when the input data is inaccurate and incomplete,the under-standing on shapes acts as the constraints to optimize the target models.For example,the geometric and topological information is useful to complete the shape and reduce the data redundancy,the semantic information helps to ensure the shape to meet the design requirement(the table surface should be planar,not spherical),the structural in-formation between components of an object or objects of an indoor scene is helpful to arrange their positions(the chair should be put in front of the desk).This paper focuses on the understanding-based 3D modeling,covering the influ-ence of different kinds of understanding information in the 3D modeling process.D-ifferent models have different properties,which call for the information of shape,se-mantics,structural relationship to improve the quality of models.The research works include:1.For object models with free shapes,we propose the medial-axis-based sparse RBF surface reconstruction algorithm.Implicit RBF surface can be used in reconstruction from the input point cloud.Re-lated works use large amount of basis functions to approximate the surface,making it inappropriate for online transmission and real-time rendering.The proposed algorithm distributes RBF centers on the medial axis of the surface,and uses sparse optimization to further reduce the number of basis functions.Medial axis,which is a shape under-standing information,helps in reducing the data redundancy and improving the quality of the models.The resulted sparse RBF surface needs much less data storage while preserving the approximation accuracy,making it suitable for the 3D modeling and the following process.2.For man-made objects,we propose the fine-grained class-aware shape dissimi-larity metric.The semantic classes of 3D models constrain their possible shapes.However,the models belonging to the same high-level class still have large variation among their shapes,which brings difficulty for shape retrieval,modeling and generation.We pro-posed the fine-grained class-aware shape dissimilarity metric algorithm,which is to first learn the latent representations of shapes using deep neural network and then dynam-ically fuse multiple fine-grained dissimilarity metrics.The experiments show that the fused metric can compare shapes based on their fine-grained semantic classes,and is robust to the overlapping between the classes.The fused shape dissimilarity metric exhibits semantic properties in shape retrieval and classification,which is helpful for novice users to retrieve target models from large database.It helps the users to acquire the proper models efficiently or re-organize the components of retrieved similar objects to obtain new models.3.For indoor scenes,we propose an efficient generation algorithm using generative recursive neural networkThe quality of indoor scenes is shown in the placements of the objects,whose plau-sibility is decided by whether the placements meet the relationship between objects.We organize the indoor scenes as hierarchical structures,and use generative recursive neu-ral network to learn the object relationship among the indoor scenes from the training data.The trained network is able to generate the corresponding plausible scene from a randomly sampled vector.We analyze the influence of different structures and object position representations by experiments,showing that the explicit structural relationship helps in indoor scene generation.