数据驱动的室内场景着色方法研究

发布时间：2018-07-03 11:30

本文选题：着色 + 室内设计　；参考：《南京大学》2017年硕士论文

【摘要】：针对一个由多个家具模型组成的室内场景,目前关于模型的布局摆放等问题的研究已经相对成熟,但关于家具模型以及场景的自动着色还没有很好的进展。而一个场景给人最直观的感觉就来源于颜色及其搭配,因此模型着色对于构建一个美观和谐的三维场景而言是至关重要的。但是对于一般人而言要选择每个物体的颜色是一个繁琐且耗时的任务,而且也难以选择出完美的颜色搭配,使得整个场景的颜色风格看起来协调一致。就算是专业的室内装饰设计师或艺术家可以根据其丰富的经验以及直觉来选择配色,为模型逐个上色也是一件麻烦的事情。如果能够实现一套为三维模型场景自动着色的方法,则不仅可以用来模拟生成仿真的室内场景,为装饰设计提供参考,也可以用于游戏场景构建,现实场景模拟等图形方面的应用。对于一个场景的着色,除了对每个物体进行着色,还要注意整个场景中物体之间颜色的搭配是否和谐,实际上对于一个物体选定其材质容易,但进一步选定材质对应的颜色却有多种选择。而另一个难题是许多模型并非是结构化的,虽然模型可以划分为一组拓扑独立的组件,但并没有语义信息,而且组件过多且复杂,对模型组件的颜色选取不能仅根据其功能性。所以颜色的选取和模型的分割是两个主要需要解决的问题。本文提出了一种基于数据驱动的方式为室内场景自动着色的方法,该方法包含以下几个方面:首先我们建立了一个数据库,包括图片-模型数据库和材质数据库两部分。我们从网络中收集了大量由设计师设计或摄影师拍摄出来的室内场景图像,以及一些家具物体的三维模型。为了建立图片与模型之间的对应关系,我们设计了一个用户友好的图片标识软件,可以更方便地完成图片中家具物体及其组件标识的工作,从而将标识细化到每个物体的组件部分,使得模型与图片中的物体在组件级别对应起来,构建了一个包括场景层,物体层和组件层三个层次的层级式图片-模型数据库。我们利用了一种新的材质表面信息获取方法构建了一个高质量的材质库,该方法结合多组拍照技术来提取材质表面的信息,为我们渲染出逼真的室内场景提供了基础。其次本文实现了对单个家具模型的着色。输入为一个三维家具物体模型和一张相同类型家具的参考图片,根据我们上一步骤构建的层级式图片-模型数据库,我们可以直观地找到图片与模型之间的对应关系,并以图片的标识信息作为参考来对家具模型进行分割。我们提取了一系列网格级和组件级的特征,对每一种分割方式进行学习从而得到图片引导的模型分割分类器,来对输入模型进行自动分割。通过将分割结果与图片中物体的不同颜色部分对应起来,我们在材质数据库中搜索与图片中家具物体的各个部分颜色最相近的材质,就可以容易地为三维模型赋予与图片相似的着色风格。最后本文实现了对场景的着色。我们提出了两种着色方式:参考示例图片或颜色主题进行着色。如果以示例图片作为参考,则可以按照对单个家具模型着色的方法为场景的中的每个物体逐个进行着色即可。如果以颜色主题作为参考,我们首先根据着色问题构建出一个马尔科夫随机场模型,并利用数据库中的数据为每个家具物体构建颜色概率分布,然后以用户的输入作为约束条件构建了一个能量方程,通过求解该方程得到每个家具模型的颜色主题,最终生成出与用户期望的颜色主题一致的室内场景着色结果。我们的实验结果以及用户调查表明我们的系统能够生成与室内设计师设计的结果相媲美的着色方案。
[Abstract]:For an indoor scene consisting of multiple furniture models, the research on the layout and placement of the model is relatively mature, but there is no good progress in the furniture model and the automatic coloring of the scene. The most intuitive feeling of a scene is from color and collocation, so the model coloring is built. It is essential for a beautiful and harmonious three-dimensional scene. But for the average person to choose the color of each object is a tedious and time-consuming task, and it is difficult to choose a perfect color match, making the color style of the whole scene look consistent. Even a professional interior decorator or art The family can choose color matching according to its rich experience and intuition. It is also a troublesome thing to color model one by one. If a set of automatic coloring methods for 3D model scenes can be realized, it can not only simulate the indoor scene of generating simulation, provide reference for decoration design, but also can be used for game scene construction. The application of real scene simulation. For a scene coloring, in addition to the color of each object, we should also pay attention to the color matching between objects in the whole scene. In fact, it is easy to select the material for one object, but there are many choices for the selection of the material for the color. And the other problem is the other problem. Many models are not structured. Although the model can be divided into a set of topology independent components, but there is no semantic information, and the components are too many and complex, the color selection of model components can not only be based on its functionality. So the selection of color and the segmentation of the model are two main problems to be solved. The method of auto coloring indoor scene based on data driven method includes the following aspects: first we set up a database, which includes two parts: picture model database and material database. We collect a large number of indoor scene images from designer or photographers from the network, and one of them. In order to establish the correspondence between the picture and the model, we have designed a user friendly image identification software, which can more conveniently complete the work of the furniture and its component identification in the picture, thus refining the identification to the component part of each object, making the model and the object in the picture in the group. A hierarchical picture model database, which includes three levels of scene layer, object layer and component layer, is constructed. We use a new material surface information acquisition method to build a high quality material library. This method combines multi group photo technology to extract information of material surface, which is rendered to us. This paper provides a basis for a realistic indoor scene. Secondly, this article implements the coloring of a single furniture model. Input into a three-dimensional furniture object model and a reference picture of the same type of furniture. According to the hierarchical picture model database built in the previous step, we can find the correspondence between the picture and the model directly. We divide the furniture model with the identification information of the picture. We extract the features of a series of grid and component levels, learn each method of the segmentation and get the model segmentation classifier which is guided by the picture to automatically divide the input model. In the color part, we search for the most similar colors of each part of the furniture in the material database. We can easily give the 3D model a coloring style similar to the picture. Finally, we make the color of the scene. We put forward two coloring methods: refer to the example picture or color theme. Line coloring. If a sample picture is used as a reference, you can coloring each object in the scene for a single furniture model. If the color topic is used as a reference, we first build a Markov random field model based on the coloring problem, and use the data in the database for each home. It constructs a color probability distribution with an object, then constructs an energy equation with the input of the user as a constraint. By solving the equation, the color theme of each furniture model is obtained, and the results of indoor scene coloring that are consistent with the desired color theme of the user are generated. Our experimental results and the user survey show that we are The system can generate a color scheme that is comparable to the result of interior designer design.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TS664.0;TU238.2

【相似文献】