The working principle of our system

我们系统的工作原理是：给定密集的照明方向，在数据捕获范围内的像素位置很有可能观察到直接从观察到的表面点上反射出来的镜面反射高光，请参见图1。来自具有已知参考几何的镜面对象的表面法线。乍看之下，给定一个像素上密集的高光集合，可以通过示例[5]（在本例中，每个像素接收从已知几何图形传输的法线）直接应用定向一致性，类似于在光度立体中进行的操作。正常恢复的全取”或阈值设置[2]将解决我们的重建问题。但是，事实证明这个问题具有挑战性，因为我们的低成本捕获设备（图2）类似于[2]，是一款动态范围有限的现成DV摄像机，在该摄像机中，由直接或间接反射引起的高光可能会以同样高的强度记录下来。为了使问题更容易解决，我们建议使用“轮廓形状”给出的法线提示，并在单个视图上标记稀疏的法线提示以跟踪真实和拒绝虚假的高光。技术贡献包括这两个法线提示的最佳集成，以得出焦油获取法线图。事实证明，可以将优化问题映射到类似于图像分割的问题，从而可以将其表达为图割优化。另一方面，我们的图形与传统的图形切割公式不同：它是双层图形，因为我们可以观察到透明对象的前景以及它背后的背景。

The working principle of our system is: given dense lighting directions, a pixel location within the data capture range has a high probability to observe a specular highlight directly reflected off the surface point being observed, see Figure 1. This can be used to obtain the surface normal from a specular object with known reference geometry. At first glance, given a dense collection of highlights at a pixel, direct application of orientation consistency similarly done in photometric stereo by examples [5] (where in our case each pixel receives a normal transferred from a known geometry) in a “winner-takes-all” or thresholded setting [2] for normal recovery would have solved our reconstruction problem. But, this problem proves to be challenging, as our low-cost capture device (Figure 2), which is similar to [2], is an off-the-shelf DV camera of limited dynamic range, where highlights caused by direct or indirect reflections are likely be recorded with equally high intensities. To make the problem tractable, we propose to use normal cues given by shape-from-silhouette, and sparse normal cues marked on a single view for tracking true and rejecting false highlights. The technical contribution consists of the optimal integration of these two normal cues for deriving the target normal map. It turns out that the optimization problem can be mapped into one similar to image segmentation and thus formulated into a graph-cut optimization. Our graph, on the other hand, is different from those in conventional graph-cut formulation: it is a dual-layered graph because we can observe a transparent object’s foreground as well as the background behind it.<br>