Research Of Interactive 3D Modeling Based On Single Images

With the rapid development of virtual reality and 3D printing technology, 3D model is widely used in different fields such as the scene modeling, robot navigation and protection of cultural heritage. Many of these applications need to obtain a 3D model from the real world. Due to complicated operation and low efficiency, traditional 3D modeling methods, including methods that based on modeling software or scene modeling based on depth information that captured from distance measuring equipment, have been unable to meet the growing demand for modeling applications. Therefore, how to build a 3D model quickly and directly based on the real scene is an important issue currently.As a pervasive medium, images are the important resources for 3D modeling that can rapidly capture the world around us. The image-based modeling techniques characterized by data acquisition convenient and photorealistic effects, have been a hot topic in computer graphics and computer vision fields.Images are lacking of the depth information about the underlying 3D scene cause of the impact of perspective projection. In the absence of the prior knowledge, it’s still a grand challenge to construct the 3D model of a target from single images.Through the observation of man-made objects in the scene, the objects, generally, can be approximately represented by a set of geometric primitives. Based on this pervasive property we presents an interactive modeling method from single images that divides the modeling problem into semantic modeling tasks and geometry tasks. Semantic tasks are achieved by the human’s high level cognitive ability, including the target object segmentation, recognition. By dragging the geometric primitives to the corresponding part of the image to achieve primitive match, the computer automatically adjust the primitive to fit the target counterpart; geometric tasks detect and infer the geometric semantic(Geo-Semantic) relations among all of the geometric primitives. In the constraints of these Geo-Semantic relationships, the computer optimize the configuration of each primitive, which is a non-convex optimization problem that requires large-scale computing.The modeling method in this paper combines the human’s cognitive ability and computer’s advantages in computing. Based on a little user’s assistance interaction, it can quickly construct a 3D photorealistic model from an uncalibrated single image. Experimental results show that 3D models constructed by geometric primitives comprise the configuration information among the various components, as well as possess the characteristics such as lightweight, ease of storage, and free for editing in the subsequent stage to applied in different context.