The following classical works deter

以下经典作品通过镜面反射确定透明对象的形状。在镜面立体[1]中，使用了两机配置和图像轨迹。在[16]中开发了镜面反射理论，其中<br>研究了镜面几何形状与图像轨迹之间的关系，并将特征分类为真实和虚拟。虚拟特征是镜面反射的反射，不限于高光，它包含有关对象形状的有用信息。在[11]中，通过利用光学现象，从物体表面反射的光的偏振度取决于反射角，而反射角又取决于表面，这两个视图用于对透明物体的表面进行建模。正常。在[10]中使用一台摄像机进一步探索了这种利用光偏振的方法，其中光的传输路径是光线追踪的。在[8]中，通过研究光传输路径，开发了关于折射和镜面3D形状的理论，只能经历两次折射。动态折射立体声使用了两种视图[12]，其中引入了折射差异的概念。在这项工作中，使用了参考图案和折射液。然后，提出了散射轨迹摄影[13]，以通过使用校准良好的捕获设备将直接散射轨迹与其他间接光学观测区分开来重建由不均匀材料制成的透明父对象。在[6]中，将透明物体浸入荧光液中。可以使用计算机控制的激光投影仪和多视图扫描进行合并。虽然在[2]中恢复的细观结构的法线图看起来不错，并被证明可用于重新照明，<br>内部反射和焦散）。我们的形状重构方法利用了粗糙的初始形状（法线），稀疏的法线提示和密集的镜面高光，这使其与上述基于光传输物理原理开发数学理论<br>或简化假设的方法脱颖而出。透明对象。

The following classical works determine the shape of a transparent object by specularities. In specular stereo [1] a two-camera configuration and image trajectory were used. A theory of specular surface was developed in [16], wherethe relationship between specular surface geometry and image trajectories were studied and features were classified asreal and virtual. Virtual features, which are reflections by a specular surface not limited to highlights, contain useful information on the shape of the object. In [11], two views were used to model the surface of a transparent object, by making use of the optical phenomenon that the degree of polarization of the light reflected from the object surface depends on the reflection angle, which in turn depends on surface normal. This approach utilizing light polarization, where the light transport paths were ray-traced, was further explored in [10] where one camera was used. In [8], a theory was developed on refractive and specular 3D shape by studying the light transport paths, which are restricted to undergo no more than two refractions. Two views were used for dynamic refraction stereo [12] where the notion of refractive disparity was introduced. In this work, a reference pattern and refractive liquid were used. Scatter trace photography [13] was then proposed to reconstruct transparent objects made of inhomogeneous materials, by using a well-calibrated capture device to distinguish direct scatter trace from other indirect optical observations. In [6], a transparent object was immersed into fluorescent liquid. Computer-controlled laser projector and multiview scans were available for merging. While the recovered normal maps of mesostructures look good in [2] and were demonstrated to be useful for relighting, their assumption on specular highlight does not apply to general transparent objects that exhibit complex refraction phenomena (such as totalinternal reflection and caustics). Our shape reconstruction method makes use of rough initial shape (normals), sparse normal cues, and dense specular highlights, which sets itself apart from the above approaches where mathematical theories were developed based on the physics of light transport,or simplifying assumptions were imposed on the transparent object.<br>