Kinect-based Data-driven 3D Modeling


Yusong Gong, Yan Zhang, Yan Wen, Zhengxing Sun

Accepted to Journal of Computer-Aided Design & Computer Graphics



This paper presents a fast modeling method based on the highly noisy and incomplete scanned data from Kinect in a single view .We take full advantage of the abundant models of the same semantic class on the Internet to build a heuristic model based on the 3D point clouds and the corresponding RGB images .Our method includes three major phases :Firstly ,we analyze the structure of the 3D point clouds and the RGB images based on the semantic segmentation of the candidate model to label the components of the scanned target ;secondly ,we search for candidate parts for the labeled parts in the target to get matched component information ;lastly ,we assemble the candidate parts to create the final 3D model .We demonstrate the efficiency and speed of our method in building models based on incomplete point clouds data from Kinect .

Overview


An overview of our approach. With an input indoor scene image, the output 3D model is reconstructed with a three-stage process: data analysis and parts retrieval and composition.





Other Scenes


Our modeling results for other objects.





PAPER: Model-driven Indoor Scenes Modeling from a Single Image