Research On 3D Object Extraction From Single Image

With the fast development of software technology as well as the remarkable improvements of hardware performance, modern computers are more and more capable of rendering large-scale 3D scenes containing numerous 3D models. This is especially true for emerging popular 3D computer games and web-based 3D virtual roaming systems, which have gained their popularity in the past decade. As a consequence, it is of great practical and theoretical significance to devise appropriate automatic or semi-automatic 3D model construction mechanisms in order to support large-scale 3D modeling.The direct geometric modeling of 3D objects requires a certain level of expertise in this specific area. Typically a model needs to be built up, modified and carefully textured before it can be utilized in applications. The whole modeling process of a single object is both time-consuming and very costly. If 3D models could be automatically(or semi-automatically with little human effort) obtained from static 2D images, we could not only attain the goal of efficiently constructing 3D models in large quantities, but also enable amateur users to easily model 3D objects. In fact, the problem of image-based 3D reconstruction is ubiquitous in the area of computer graphics and computer-aided geometric design and a large body of work has already been carried out. Through tackling this problem, we provide a feasible idea and vision for the ultimate objective of large-scale automatic 3D modeling.Concretely, our main contributions are listed as follows.1. Based on existing approaches for image-based 3D modeling, we study the possibility of using traditional computer-aided design methods, namely sweeping, lofting and revolving in the context of 3D object extraction from single image. In the meantime, we exploit several other methods such as hierarchical edge feature extraction, 8-neighborhood point linkage and quasi-Newton method for real-time profile snapping in our approach. Compared with existing methods for single-image-based 3D reconstruction, our system supports a considerably larger range of 3D parts.2. Our system provides users with the possibility of dynamically editing 3D parts through scaling, rotating and transitioning. Also, our system supports composition of multiple 3D parts to construct a relatively complex 3D model.3. We compare our work with similar existing approaches in various aspects and discuss the feasibility of using automatic scene depth estimation techniques in our problem context.To summarize, through using our 3D modeling tool, amateur users can also efficiently and elegantly extract complete 3D models from images, making them fit with their counter-parts in these images. They can further edit and use the extracted models in the way they want. This provides a feasible realization path for the ultimate goal of large-scale automatic 3D modeling. For example, our system can be integrated into emerging 3D systems such as online 3D roaming systems and 3D visualization systems so as to provide 3D models in large quantities for them. In the future we plan to consider supporting freer 3D parts and more nature 3D objects. Besides, since the backside of an object is invisible from only one photo, we could possibly make use of a database of stock models with complete geometry and texture information to back up the modeling process by providing lost appearances where appropriate.