Toward General and Robust Visible Surface Reconstruction

这是为期三年的连续奖项的第一年。本研究探索了一种新的计算机视觉中的表面重构方法。表面重建是从一组噪声甚至是稀疏的深度点中恢复场景中可见表面的描述的过程，在物体建模、抓取规划和导航等机器人任务中起着至关重要的作用。标准方法有两个局限性：它们假设数据中没有极端误差；它们假设可以用图像坐标的单个分段连续函数精确地模拟可见表面。这些局限性严重制约了现有曲面重建技术的适用性。新方法使用两阶段技术克服了这些问题：(1)局部估计阶段，在整个图像区域产生多项式曲面片，可能包括重叠的表面；(2)全局细化阶段，链接和细化相互一致的局部估计。

This is the first year of a three-year continuing award. The research investigates a new approach to surface reeconstruction in computer vision. Surface reconstruction, the process of recovering descriptions of the visible surfaces in a scene from a noisy and perhaps sparse set of depth points, plays a fundamental role in such robotic tasks as object modeling, grasp planning and navigation. Standard approaches have two limitations: they assume that there are no extreme errors in the data; and they assume that the visible surface can be accurately modeled with a single piecewise continuous function of the image coordinates. These limitations severely restrict the applicability of present surface reconstruction techniques. The new approach overcomes these problems using a two phase technique: (1) a local estimation phase that produces polynomial surface patches in regions throughout the image, including possibly overlapping surfaces, and (2) a global refinement phase that links and refines mutually consistent local estimates.