Abstract: 3D reconstruction of real world objects is a fundamental problem in computer vision and computer graphics. It is a challenge for high fidelity reconstruction with consumer-grade depth camera, e.g. Microsoft Kinect on large scale scene. A divide-and-conquer approach could both preserve local geometry details of scene objects and global consistency of the whole scene, thus is superior to existing GPU based real-time/global optimization based offline methods. In this paper, we extend the state-of-the-art in two folds. First, more fine grit scene segmentation is conducted for better local details preservation. Second, a scene analysis procedure is proposed for more reasonable global error spread. Preliminary experiments show that our method achieves more convincing reconstruction results.