Research On Data-driven Approach For Furniture And Indoor Scene Colorization

For an interior scene with multiple furniture models,there have been progres-sive improvements to help users to create effective furniture layout in recent years,but colorization for furniture objects and further the whole indoor scene has not yet re-ceived the considerable attention that it deserves.The fact is that much of what people perceive and feel about the scene is experienced through colors and their combination.Therefore,colorization of models does matter for building a visually pleasing 3d scene.However,this is a tedious and daunting task to choose the color of every object;it is also difficult to combine the colors of objects to make the whole scene harmonious.Even for experienced interior designer or artist,assigning color to models one by one is fussy.So if we can find a way to colorize 3D scene automatically,not only can we use it for interior design,but also for other applications such as game scene and realism.To colorize a scene,except assigning color to every object,the combination of all the objects in the scene is also important.In fact determining the texture of an object is easy,but there are many colors can be chosen for one texture.Another problem is that furniture models are often unstructured.Although the furniture model is composed of a set of topologically independent components,each of which is not necessarily to be semantically meaningful.Even though semantic segmentation applies to the model,the object is often not colorized completely according to different functional parts.So the main two problems is the selection of color and segmentation of models.Here we present a data-driven approach that colorizes 3D furniture models and indoor scenes automatically,including following steps:first of all,we build a database, including a image-model database and a texture database.We collect a large quantity of works by professional interior designers,also some 3D furniture models from the internet.To build correspondences between images and models,we develop a user-friendly prototype,making the task of furniture and components labelling easier.With the labelling data,we build a well-structured image-model database with detailed an-notations and image-model correspondences,including scene level,object level and part level.By using multiple photography skills to extract textures’ surface informa-tion,we also build a texture database with high quality,which is the base for rendering realistic results.Furthermore,we implement colorization for single furniture model.With the in-put of a 3D furniture model and a same type furniture image as reference,according to the database,we can find correspondence between images and furniture models di-rectly.By using color and segmentation information of furniture in images and a serials extracted features at mesh and part level,we can train the classifier with multiple seg-mentation schemes.After matching the different parts in images and models,we can colorize the furniture easily by searching textures with similar colors.Finally yet importantly,we implement colorizing whole 3D scene.We provide two circumstances:referring image or color theme.For referring image,we can col-orize every model using the approach for colorizing single furniture model.For refer-ring color theme,we first build a Markov Random Field model and color probability distribution for objects of every type,then formulate an energy function with user in-put as constraint.By optimizing the function,we get color theme for every model and finally colorize the 3D indoor scene resembling user desired color theme.Our experi-ments and a user study show that our system produces perceptually convincing results comparable to those generated by interior designers.