photometric method for determining surface orientation from multiple images-读书笔记

从多个图像确定表面方向的光度法
photometric method for determining surface orientation from multiple images

Abstract
photometric stereo:
vary the direction of incident illumination between successive images, while holding the viewing direction constant.
改变连续图像之间的入射光方向，同时保持观察方向不变
this provides sufficient information to determine surface orientation at each image point.
这为确定每个图像点的表面方向提供了足够的信息
Since the imaging geometry is not changed, the correspondence between image points is known a priori.
由于成像几何没有改变，图像点之间的对应关系是先验的

该技术是光度学的，因为它使用的是在连续视图中单个图像位置记录的辐射值，而不是在基于计算机的图像理解中使用的置换特征的相对位置光度学立体。
它有两种应用方式：
首先，它是确定每个图像点的表面方向的通用技术。
第二，它是一种确定具有特定表面方向的目标点的技术，这些应用通过综合实例加以说明。
关键词：双向反射分布函数BRDF图像处理、成像几何、入射光光度立体、反射图、表面方向

1. INTRODUCTION
   图像的投影几何the geometry of image projection
   影像形成的辐射测量学 the radiometry of image formation
   Relating the radiance values recorded in an image to object shape requires a model of the way surfaces reflect light .
   将图像中记录的辐射值与物体形状相关联需要一个表面反射光线的方式的模型

A reflectance map is a convenient way to incorporate a fixed scene illumination, surface reflectance and imaging geometry into a single model that allows image intensity to be written as a function of surface orientation. This function is not invertible since surface orientation has two degrees of freedom and image intensity provides only one measurement. Local surface shape cannot, in general ,be determined from the intensity value recorded at a single image point. In order to determine object shape, additional information must be provided.
反射贴图是将固定场景照明、表面反射和成像几何体合并到单个模型中的一种方便方法，该模型允许图像强度作为表面方向的函数写入。此函数不可逆，因为表面方向有两个自由度，图像强度仅提供一个测量。通常，局部表面形状不能从单个图像点记录的强度值确定。为了确定对象形状，必须提供其他信息。

This observation has led to a novel technique called photometric stereo in which surface orientation is determined from two or more images.
这种观察已经导致了一种称为光度立体的新技术，在这种技术中，表面方向是由两个或多个图像决定的。
Traditional stereo techniques determine range by relating two images of an object viewed from different directions. If the correspondence between picture elements is known, then distance to the object can be calculated by triangulation. Unfortunately, it is difficult to determine this correspondence.
The idea of photometric stereo is to vary the direction of the incident illumination between successive images, while holding the viewing direction constant. It is shown that this provides sufficient information to determine surface orientation at each image point. Since the imaging geometry is not changed, the correspondence between image points is known a priori. The technique is photometric because it uses the radiance values recorded at a single image location, in successive views, rather than the relative positions of displaced features.
传统的立体技术通过关联从不同方向观看的物体的两个图像来确定距离。如果知道图像元素之间的对应关系，则可以通过三角测量计算到对象的距离。不幸的是，很难确定这种对应关系。
光度立体的概念是改变连续图像之间的入射光的方向，同时保持观察方向不变。结果表明，这为确定每个图像点的表面方向提供了足够的信息。由于成像几何没有改变，图像点之间的对应关系是先验的。该技术是光度学的，因为它使用在连续视图中的单个图像位置记录的辐射值，而不是替换特征的相对位置。

2. THE REFLECTANCE MAP
   The fraction of light reflected by an object surface in a given direction depends upon the optical properties of the surface material, the surface microstructure and the spatial and spectral distribution and state of polarization of the incident illumination. For many surfaces, the fraction of the incident illumination reflected in a particular direction depends only on the surface orientation. The reflectance characteristics of such a surface can be represented as a function (i,e,g)of the three angles i, e and g , defined in Figure 1. These are called, respectively, the incident , emergent and phase angles. The angles i and e are defined relative to a local surface normal. /(i, e, g)determines the ratio of surface radiance to irradiance measured per unit surface area, per unit solid angle, in the direction of the viewer. The reflectance function * (i, e, g) defined here is related to the bidirectional reflectance distribution function (BRDF) defined by the National Bureau of Standards.
   反射率图
   物体表面在给定方向上反射光的比例取决于表面材料的光学性质,表面微观结构和入射光的空间光谱分布及偏振状态。对于许多曲面，在特定方向上反射的入射光的比例仅取决于曲面方向。这种表面的反射特性可以表示为图1中定义的三个角（即和）的函数（即，g）。它们分别称为入射角、出射角和相位角。角度i和e是相对于局部曲面法向定义的。/（i，e，g）确定每单位表面积、每单位立体角测量的表面辐射与辐照度之比。在观察者的方向。这里定义的反射函数*（即，g）与国家标准局定义的双向反射分布函数（BRDF）有关。

Image forming systems perform a perspective transformation, as illustrated in Figure 2 (a). If the size of the objects in view is small compared to the viewing distance, then the perspective projection can be approximated as an orthographic projection, as illustrated in Figure 2(b). Consider an image forming system that performs an orthographic projection. To standardize the imaging geometry, it is convenient to choose a coordinate system such that the viewing direction is aligned with the negative z-axis. Also, assume appropriate scaling of the image plane such that object point(x, y, z)maps onto image point(u,])where u=x and v=y With these assumptions, image coordinates(x, y)and object coordinates (x, y)can be referred to interchangeably.
If the equation of an object surface is given explicitly as:

then a surface normal is given by the vector:

If parameters p and q are defined by:

图像形成系统执行透视变换，如图2（a）所示。

如果与视距相比，视图中对象的大小较小，则透视投影可以近似为正交投影，如图2（b）所示。

考虑一个执行正交投影的图像形成系统。
为了使成像几何图形标准化，可以方便地选择一个坐标系，使观察方向与负Z轴对齐。此外，假设图像平面的适当缩放，使对象点（x，y，z）映射到图像点（u，]），其中u=x和v=y在这些假设下，图像坐标（x，y）和对象坐标（x，y）可以互换引用。
如果物体表面的方程式明确地给出如下：
然后由向量给出曲面法向：
then the surface normal can be written as [p, q, -1]. The quantity (p, g) is called the gradient of f(x,y) and gradient space is the two-dimensional space of all such points(p, g). Gradient space is a convenient way to represent surface orientation. It has been used in scene analysis. In image analysis, it is used to relate the geometry of image projection to the radiometry of image formation. This relation is established by showing that image intensity can be written explicitly as a function of gradient coordinates p and q .
那么曲面法向可以写成[P，Q，-1]。量（p，g）称为f（x，y）的梯度，梯度空间是所有这些点（p，g）的二维空间。梯度空间是表示曲面方向的一种简便方法。它已用于场景分析。在图像分析中，它用于将图像投影的几何结构与图像形成的辐射测量联系起来。这一关系的建立表明，图像强度可以明确地表示为梯度坐标p和q的函数。
An ideal imaging device produces image irradiances proportional to scene radiances. In an orthographic projection, the viewing direction, and hence the phase angle g, is constant for all object points.
Thus, for a fixed light source and viewer geometry, the ratio of scene radiance to irradiance depends only on gradient coordinates p and q. Further, suppose each object surface element receives the same incident radiance. Then, the scene radiance and hence image intensity, depends only on gradient coordinates p and q .
理想的成像设备产生与场景辐射成比例的图像辐射。在正交投影中，观察方向，因此相位角G，对于所有对象点都是恒定的。
因此，对于固定光源和观察者几何体，场景辐射与辐照度之比仅取决于梯度坐标p和q。此外，假设每个物体表面元素接收相同的入射辐射。然后，场景辐射和图像强度仅取决于梯度坐标p和q。
The reflectance map R(p, g)determines image intensity as a function of p and q. A reflectance map captures the surface reflectance of an object material for a particular light source, object surface and viewer geometry. Reflectance maps can be determined empirically, derived from phenomenological models of surface reflectivity or derived from analytic models of surface microstructure.
反射比图r（p，g）确定图像强度作为p和q的函数。反射比图捕获特定光源、对象表面和查看器几何体的对象材质的表面反射比。反射率图可以通过经验确定，可以从表面反射率的现象学模型推导，也可以从表面微观结构的分析模型推导。
In this paper, it will be assumed that image projection is orthographic and that incident illumination is given by a single distant point source. Extended sources can be modeled as the superposition of single sources. The reflectance map can be extended to incorporate spatially varying irradiance and perspective. A formal analysis of the relation between the reflectance map and the bidirectional reflectance distribution function(BRDF)has been given.
Expressions for cos( ) cos(e)and cos(g)can be derived using normalized dot products of the surface normal vector [p, q, -1], the vector [ps,@s, -1] which points in the direction of the light source and the vector [0,0, -1] which points in the direction of the viewer One obtains:
本文假设图像投影是正交的，入射光由单个远点光源给出。扩展源可以建模为单个源的叠加。反射图可以扩展到包含空间变化的辐照度和透视图。对反射图与双向反射分布函数（BRDF）的关系进行了形式化分析。

cos（i）cos（e）和cos（g）的表达式可以使用表面法向量[p，q，-1]的归一化点积、指向光源方向的向量[ps，qs，-1]和指向观察者方向的向量[0,0，-1]得出：

These expressions can be used to transform an arbitrary surface reflectance function *(i,e, g)into a reflectance map R(p,q).
One simple idealized model of surface reflectance is given by

This reflectance function corresponds to the phenomenology model of a perfectly diffuse( lambertian )surface which appears equally bright from all viewing directions. Here,Q is a reflectance factor and the cosine of the incident angle accounts for the foreshortening of the surface as seen from the source. The corresponding reflectance map is given by:

A second reflectance function, similar to that of materials in the maria of the moon and rocky planets, is given by:

这些表达式可用于将任意表面反射率函数*（即，g）转换为反射率图r（p，q）。
表面反射率的一个简单理想化模型由下式给出：
这个反射函数对应于一个完全漫反射（朗伯）表面的现象学模型，它在所有观察方向上看起来都一样明亮。这里，q是一个反射系数，入射角的余弦值解释了从源头看表面的缩短。相应的反射率图由下式给出：
第二个反射函数，类似于月球和岩石行星玛丽亚的物质反射函数，由以下公式给出：

This reflectance function corresponds to the phenomenological model of a surface which reflects equal amounts of light in all directions. The cosine of the emergent angle accounts for the foreshortening of the surface as seen from the viewer. The corresponding reflectance map is given by:

It is convenient to represent R(p, q)as a series of iso-brightness contours in gradient space. Figure 3 and Figure 4 illustrate the two simple reflectance maps Ra(P g)and Rb(p, g), defined above for the case ps 0.7, qs =0.3 and e= 1
这个反射函数对应于一个表面的现象学模型，它在所有方向上反射等量的光。从观察者的角度看，出射角的余弦表示表面的缩短(按照透视法)。相应的反射率图由下式给出：
将R（P，Q）表示为梯度空间中的一系列等亮度等值线是方便的。图3和图4说明了两个简单的反射比图Ra（p g）和Rb（p，g），上面定义的情况为ps 0.7，qs=0.3和e=1。

Figure 3. The reflectance map R(p,q) for a lambertian surface illuminated from gradient point ps 0.7 and qs 0.3 (with e-1.0). The reflectance map is plotted as a series of contours spaced 0.1 units
Figure 4. The reflectance map R(p,q)for a surface illuminated from gradient point p = 0.7 and a,= 0. 3(with o=1.0). The reflectance map is plotted as a series of contours spaced 0.2 units apart.
图3从提督店 PS 0.7 and QS 0.3（with Q=1.0）中的反射图（P，Q）。反射比图以一系列间隔0.1个单位的等高线绘制。
图4从梯度点p = 0.7和A，= 0的反射图映射图（p，q）。3（with Q = 1.0）。反射比图绘制成一系列相隔0.2个单位的等高线。
Reflectance map techniques
Using the reflectance map, the basic equation describing the Image-forming process can be written as：
I(x,y)=R(p, q) (1)
One idea is to use eq. (1 ) directly to generate shaded images of surfaces, this has obvious utility in computer graphics applications including hill-shading for automated cartography ’ and video input for a flight simulator . Synthesized imagery can be registered to real imagery to align images with surface models this technique has been used to achieve precise alignment of Landsat imagery with digital terrain models.
反射图技术
利用反射率图，描述图像形成过程的基本方程可以写成：
i（x，y）=r（p，q）（1）
一种想法是直接使用等式（1）生成表面的阴影图像，这在计算机图形应用中具有明显的实用性，包括自动制图的小山阴影和飞行模拟器的视频输入。合成图像可以注册到真实图像中，以使图像与地面模型对齐。该技术已用于实现陆地卫星图像与数字地形模型的精确对齐。
equation (1) can also be used in image analysis to determine object shape from image intensity .equation (1) is a nonlinear first order partial differential equation , direct solution is tedious .More generally , one can think of eq (1 ) as one equation in the two unknowns p and q. Determining object shape from image n -tension is difficult because eq . ( 1) is underdetermined . in order to calculate object shape additional assumptions must be invoked .
Recent work has helped to make these assumptions explicit. For certain materials such as the material of the maria of the moon special properties of surface reflectance simply the solution .Other methods for determining object shape from image intensity embody assumptions about surface curvature, 10 simple surfaces have been proposed for use in computer aided design . when properties of surface curvature are known a priori , they can be exploited in image analysis . this is useful ,for example , in industrial inspection since there are often constraints on surface curvature imposed by the drafting techniques available for part design and by the fabrication processes available for part manufacture .
Reflectance map techniques deepen our understanding of what can and cannot be computed directly from image intensity. Photometric stereo is a novel reflectance map technique that uses two or more images to solve eq. (1 ) directly
方程（1）也可用于图像分析中，根据图像强度确定物体形状。方程（1）是一个非线性一阶偏微分方程，直接解是繁琐的。一般来说，可以将方程（1）看作是两个未知数p和q中的一个方程。根据图像n-张力确定物体形状是困难的，因为（1）不确定。为了计算对象形状，必须调用其他假设。
最近的工作有助于明确这些假设。对于某些材料，如月球的材料，表面反射的特殊性质，简单地解决了这个问题。其它根据图像强度确定物体形状的方法，包含了表面曲率的假设，提出了10个简单表面，用于计算机辅助设计。当曲面曲率的性质已知时，可以在图像分析中加以利用。例如，这在工业检验中很有用，因为零件设计所用的绘图技术和零件制造所用的制造工艺通常会对表面曲率施加限制。
反射比地图技术加深了我们对直接从图像强度计算什么和不能直接从图像强度计算什么的理解。光度立体是一种新的反射率地图技术，它使用两个或多个图像直接求解等式（1）。

III PHOTOMETRIC STEREO
The idea of photometric stereo is to vary the direction of incident illumination between successive views, while holding the viewing direction constant. Suppose two images I1(x, y)and I2(x, y)are obtained by varying the direction of incident illumination. Since there has been no change in the imaging geometry, each picture element (x, y) in the two images corresponds to the same object point and hence to the same gradient (p, q). The effect of varying the direction of incident illumination is to change the reflectance map R (p, q) that characterizes the imaging situation .
Let the reflectance maps corresponding to I1(x, y) and I2(x, y) be R1 (p, q) and R2 (p, q) respectively. The two views are characterized by two independent equations:
I1(x, y) = R1 (p, q) (2)
I2(x y) = R2 (p, q) (3)
光度立体
光度立体的概念是改变连续视图之间的入射光方向，同时保持观察方向不变。假设通过改变入射光的方向得到两幅图像I1（x，y）和I2（x，y）。由于成像几何结构没有变化，两幅图像中的每个图像元素（X）对应于相同的目标点，因此对应于相同的梯度（P，Q）。改变入射光方向的效果是改变表征成像情况的反射比图R（P，Q）。
让I1（x，y）和I2（x，y）对应的反射率图分别为R1（p，q）和R2（p，q）。这两种观点的特点是有两个独立的方程式：
I1（x，y）=R1（p，q）（2）
I2（x y）=r2（p，q）（3）
Two reflectance maps R1(p ,q)and R2(p, q) are required. But, if the phase angle g is the same in both views(i.e., the direction of illumination is rotated about the viewing direction), then the two reflectance maps are rotations of each other.
For reflectance characterized by Rb (p, q)above, Eqs. (2)and(3)are linear equations in p and q. If the reflectance factor Q is known, then two views are sufficient to determine surface orientation at each image point, provided the directions of incident illumination are not collinear in azimuth.
In general, Eqs.(2)and(3)are nonlinear so that more than one solution is possible. One idea would be to obtain a third image:
I3(x,y)= R3(p, q)(4)
to overdetermine the solution.
For reflectance characterized by Ra(p, q)above, three views are sufficient to uniquely determine both the surface orientation and the reflectance factor e at each image point, as will now be shown. Let I =[l1, l2, l3] be the column vector of intensity values recorded at a point (x y) in each of three views (denotes vector transpose). Further, let
N1=(n11, n12,n13]
需要两个反射率图R1（P，Q）和R2（P，Q）。但是，如果两个视图中的相位角g相同（即，照明方向围绕观察方向旋转），则两个反射率图是彼此的旋转。
对于上面以Rb（p，q）为特征的反射率，方程。（2）和（3）是p和q中的线性方程。如果已知反射系数q，则两个视图足以确定每个图像点的表面方向，前提是入射光的方向在方位上不共线。
一般来说，方程（2）和（3）是非线性的，因此可以有多个解。一个想法是获得第三个图像：
I3（x，y）=r3（p，q）（4）
过度确定解决方案。
对于上面以ra（p，q）为特征的反射，三个视图足以唯一地确定每个图像点的表面方向和反射系数Q，如下所示。设I=[l1，l2，l3]为三个视图（表示矢量转置）中每一个点（x y）记录的强度值的列向量。进一步，让
n1=（n11，n12，n13）
n2=（n21，n22，n23）
n3=（n31，n32，n33）
be unit column vectors defining the three directions of incident illumination. Construct the matrix N where
N=
Let n (n1, n2, n3]be the column vector corresponding to a unit surface normal at (x, y). Then
I=Q N n
so that
Q n=N^(-1)*I
provided the inverse N^ (-1)exists. This inverse exists if and only if the three vectors n1 n2 and n3 do not lie in a plane. In this case, the reflectance factor and unit surface normal at (x, y) are given by:
n = (1/Q)N^(-1)*I (5)
Unfortunately, since the sun’s path across the sky is very nearly planar, this simple solution does not apply to outdoor images taken at different times during the same day.
定义入射光的三个方向的单位列向量。构造矩阵n，其中
n=
设n（n1，n2，n3]为对应于（x，y）处单位曲面法向的列向量。然后
I= q n n
以便
qn=n^（- 1）*i
如果存在倒数n^（-1）。只有当且仅当三个矢量n1 n2和n3不在平面上时，才存在该逆。在这种情况下，反射系数和单位表面法向（x，y）由以下公式给出：
n=（1/q）n^（-1）*i（5）
不幸的是，由于太阳穿过天空的路径几乎是平面的，这个简单的解决方案不适用于同一天在不同时间拍摄的室外图像。
Even when the simplifications implied by Ra(p, q) and Rb (p, q)above do not hold, photometric stereo is easily implemented. Initial computation is required to determine the reflectance map for each experimental situation. Once calibrated, however photometric stereo can be reduced to simple table lookup and/or search operations. Photometric stereo is a practical scheme for environments, such as industrial inspection, in which the nature and position of the incident illumination is known or can be con-trolled .
The multiple images required for photometric stereo can be obtained by explicitly moving a single light source, by using multiple light sources calibrated with respect to each other or by rotating the object surface and imaging hardware together to simulate the effect of moving a single light source. The equivalent of photometric stereo can also be achieved in a single view by using multiple illuminations which can be separated by color.
即使上面的ra（p，q）和rb（p，q）所暗示的简化不成立，光度立体图也很容易实现。初步计算需要确定每个实验情况下的反射比图。然而，一旦校准，光度立体可以简化为简单的表格查找和/或搜索操作。光度立体是工业检测等环境中的一种实用方案，在这种环境中，入射光的性质和位置是已知的或可以控制的。
通过明确移动单个光源、使用相互校准的多个光源或通过旋转物体表面和成像硬件来模拟移动单个光源的效果，可以获得光度立体所需的多个图像。通过使用可以用颜色分隔的多个照明，也可以在单个视图中实现等效的光度立体。

Applications of photometric stereo
Photometric stereo can be used in two ways. First, photometric stereo is a general technique for determining surface orientation at each image point. For a given image point(x, y ) the equations characterizing each image can be combined to determine the corresponding gradient(p, q)
Second, photometric stereo is a general technique for determining object points that have a particular surface orientation This use of photometric stereo corresponds to interpreting the basic image-forming Eq. (1)as one equation in the unknowns x and y. For a given gradient (p, q), the equations characterizing each image can be combined to determine corresponding object points(x, y). This second use of photometric stereo is appropriate for the so-called industrial "bin-of-parts "problem. The location in an image of key object points is often sufficient to determine the position and orientation of a known object on a table or conveyor belt so that the object may be grasped by an automatic manipulator
光度立体的应用
光度立体可以用两种方式。首先，光度立体是确定每个图像点的表面方向的通用技术。对于给定的图像点（x，y），可以将表征每个图像的方程组合起来，以确定相应的梯度（p，q）。
其次，光度立体是确定具有特定表面方向的对象点的通用技术。使用光度立体对应于将基本图像形成方程（1）解释为未知x和y中的一个方程。对于给定的梯度（p，q），可以将表征每个图像的方程组合起来确定对应的对象点（x，y）。这种第二次使用的光度立体是适合所谓的工业“零件仓”的问题。关键目标点图像中的位置通常足以确定已知目标在工作台或传送带上的位置和方向，以便自动操纵器能够抓住目标。
A particularly useful special case concerns object points whose surface normal directly faces the viewer (i.e, object points with p=0 and q-0) Such points form a unique class of image points whose intensity value is invariant under rotation of the illumination direction about the viewing direction. Object points with surface normal directly facing the viewer can be located without explicitly determining the reflectance map R(p,q). The value of R(0.0)is not changed by varying the direction of illumination provided only that the phase angle g is held constant.
These applications of photometric stereo will now be illustrated using a simple, synthesized example. Consider a sphere of radius r centered at the object space origin. The explicit representation of this object surface, corresponding to the viewing geometry of Figure 2(b), is given by:

一个特别有用的特殊情况涉及其表面法向直接面向观察者的对象点（即P=0和q=0的对象点），这些点形成一类独特的图像点，其强度值在围绕观察方向的照明方向旋转下不变。在不明确确定反射比图r（p，q）的情况下，可以定位表面垂直于观察者的对象点。只要相位角g保持不变，r（0.0）的值不会因改变照明方向而改变。
现在将用一个简单的合成例子来说明光度立体的这些应用。考虑以对象空间原点为中心的半径为r的球体。与图2（b）的观察几何体相对应的该物体表面的显式表示如下：

The gradient coordinates p and q are determined by differentiating Eq. (6) with respect to x and y. One finds:

Suppose that the sphere is made of a perfectly diffusing object material and is illuminated by a single distant point source at gradient point (ps, qs]… Then, the reflectance map is given by Ra(p, q) above so that the corresponding synthesized image is:

Equation (7) generates image intensities in the range 0 to Q. In the example below r=60 and Q= 1.
Multiple images are obtained by varying the position of the light source. Consider three different positions. Let the first be Ps=0.7 and qs=0.3 as in Figure 3. Let the second and third correspond to rotations of the light source about the view in direction of and +120 restively (i.e. ps=0.610. qs=0.456and ps =-0.090. (Is =-0.756). Let the three reflectance maps be R1 (p, q) ,R2 (p, q) and R3(p, q). The phase angle g is constant in each case. Let the corresponding images generated by Eq. (6) be I1(x, Y). I2(x, Y) and I3(x, Y) .
梯度坐标p和q是通过对等式（6）对x和y进行微分来确定的。我们发现：

假设球体由一个完全扩散的物体材料制成，并由一个位于梯度点（ps，qs）的远点光源照亮。然后，由上面的Ra（p，q）给出反射率图，因此相应的合成图像为：

方程式（7）生成的图像强度范围为0到Q。在下面的示例中，r=60和q=1。
通过改变光源的位置来获得多个图像。考虑三个不同的位置。第一个为ps=0.7，qs=0.3，如图3所示。让第二个和第三个对应于光源围绕视图沿和+120方向的旋转（即ps=0.610）。qs=0.456，ps=-0.090。（IS=-0.756）。让三个反射比图为R1 (p, q) ,R2 (p, q) and R3(p, q)。相位角g在每种情况下都是恒定的。由式（6）生成的相应图像为I1（x，y）。I2（x，y）和I3（x，y）。
First, consider image point x= 15, y=20. Here, I1(x, y) = 0.942. I2(x, y) = 0.723 and I3(x, y) = 0.505 Figure 5 illustrates thee reflectance map contours R1(p, q)=0.942, R2(p, q)=0.723 and R3(p, q) =0.505. The point p=0.275, q =0.367 at which these three contours intersect determines the gradient corresponding to image point x= 15, y =20.
Second, consider gradient point p=0.5,q=0.5.Here,R1(p, q)=0.974,R2(p, q)=0.60 and R3(p, q)=0.375. Figure (6) illustrates the image intensity contours I1 (x, y)=0.974, l2(x y)=0.600 and l3(x, y)=0.375. The point x=24.5 y=24.5 at which these three contours intersect determines an object point whose gradient is p=0.5, q=0.5
Finally, Figure 7 repeats the example given in Figure 6 but for the case p=0, q=0 . Here R1(p, q)= R2(p, q)= R3(p, q) =0.796 . Object points with surface normal directly facing the viewer form a unique class of points whose image intensity is in variant for rotations of the light source about the viewing direction . The point x =0 , y=0 at which these three contours intersect determines an object point with surface normal directly facing the viewer. This result would hold even if the form of R(p , q) is unknown.
首先，考虑图像点x=15，y=20。这里，I1（x，y）=0.942。I2（x，y）=0.723和I3（x，y）=0.505图5说明了反射率图轮廓R1（p，q）=0.942，r2（p，q）=0.723和r3（p，q）=0.505。这三条等高线相交的点P=0.275，Q=0.367决定了与图像点X=15，Y=20相对应的梯度。
其次，考虑梯度点p=0.5，q=0.5。这里，r1（p，q）=0.974，r2（p，q）=0.60，r3（p，q）=0.375。图（6）显示了图像强度轮廓I1（x，y）=0.974，l2（x y）=0.600和l3（x，y）=0.375。这三条等高线相交的点x=24.5y=24.5决定了一个物体点，其梯度为p=0.5，q=0.5。
最后，图7重复了图6中给出的示例，但对于p=0，q=0的情况。这里，r1（p，q）=r2（p，q）=r3（p，q）=0.796。表面垂直于观察者的对象点形成了一类独特的点，其图像强度随光源围绕观察方向的旋转而变化。这三条轮廓相交的点x=0，y=0决定了一个对象点，其表面法向直接面向观察者。即使R（P，Q）的形式未知，这个结果也会保持不变。

Accuracy considerations
Photometric stereo is most accurate in regions of gradient space where the density of reflectance map contours is great and where he contours to be intersected are nearly perpendicular. Several factors influence the density and direction of reflectance map contours. The reflectance properties of the surface material play a role. Figures 3 and 4 illustrate the difference between two idealized materials viewed under identical conditions of illumination ，In general, increasing the specular component of reflection will increase the density of contours in one region of gradient space at the expense of other regions. Using extended light sources rather than point sources will alter the shape and distribution of reflectance map contours. Imaging systems can be configured to exploit these facts.
For a given surface material, the main determiner of accuracy is the choice of phase angle g. In photometric stereo, there is a trade-off to acknowledge. A large phase angle increases the density of reflectance map contours in illuminated portions of gradient space. At the same time, a large phase angle results in more of gradient space lying in shadow. A practical compromise must be arrived at for each application.
精度注意事项
光度立体是梯度空间中最精确的，在梯度空间中，反射率图轮廓密度较大，且要相交的轮廓几乎垂直。影响反射率图等值线密度和方向的因素有很多。表面材料的反射特性起着一定的作用。图3和图4说明了在相同的照明条件下观察到的两种理想化材料之间的差异，一般来说，增加反射的镜面分量将增加梯度空间一个区域中轮廓的密度，而其他区域将受到影响。使用扩展光源而不是点光源将改变反射图轮廓的形状和分布。成像系统可以配置为利用这些事实。
对于给定的表面材料，精度的主要决定因素是相位角G的选择。在光度立体中，有一个权衡需要承认。较大的相位角会增加梯度空间被照明部分的反射率图轮廓密度。同时，大的相位角会导致更多的梯度空间处于阴影中。每个应用程序都必须达成一个实际的折衷方案。

The relative positions of the light sources must also be considered. Figures 8 and 9 give some indication of the trade-off associated with light source position. In each case. reflectance is assumed to be characterized by Ra(p, q)above. Figure 8 considers a two-source configuration in which the light source directions are separated by 90 in azimuth with respect to the viewer. Figure 8 superimposes reflectance map contours, spaced 0.1 units apart , for R1(p, q) and R2(p, q) where R1(p, q) is Ra(p q) with ps = 0.7, qs =0.3 ,Q=1 and R2(p, q) is Ra(p q) with ps = -0.3, qs =0.7 ,Q=1 . Each region of Figure 8 corresponds to a region of equal measurement error. For example, if I1(x, y) is determined to lie between 0.4 and 0.5 and l2 (x, y) is determined to lie between 0.5 and 0.6 then surface orientation can be determined to 68 of its true value. This corresponds to an area of gradient space in the third quadrant where the error regions are small.
还必须考虑光源的相对位置。图8和图9给出了与光源位置相关的权衡。在每种情况下。假设反射比具有上述Ra（p，q）的特征。图8考虑了两个光源配置，其中光源方向相对于观察者的方位角被90分开。图8:R1（p，q）和R2（p，q）的叠加反射率图轮廓，间隔0.1个单位，其中R1（p，q）是Ra（p q），其中ps=0.7，qs=0.3，q=1，r2（p，q）是Ra（p q），其中ps=-0.3，qs=0.7，q=1。图8的每个区域对应一个测量误差相等的区域。例如，如果确定I1（x，y）介于0.4和0.5之间，并且确定L2（x，y）介于0.5和0.6之间，则可将表面方向确定为其真实值的68。这对应于第三象限中的梯度空间区域，其中误差区域很小。
Here, a measurement error of I gray level in 10 in each of I1(x, y) and I2(x, y) constrains surface orientation to within +6.8… On the other hand, if I1(x, y) is determined to lie between 0.9 and 1.0 and I2(x, y) is determined to lie between 0.5 and 0.6 then surface orientation can be determined to 25.8 of its true value. This corresponds to an area of gradient space in the first quadrant where the error regions are large. Here, a measurement error of 1 gray level in 10 in each of I1(x, y) and I2(x, y) only constrains surface orientation to within + 25.80
Figure 9 repeats the example of Figure 8 but with the second light source separated by 180 in azimuth from the first. In this configuration, error regions are smallest in the second and fourth quadrants of gradient space. Combinations using more than two light sources can be arranged to achieve a desired overall accuracy. One idea is to choose four directions of illumination ,spaced evenly in azimuth with respect to the viewer and having a relatively large phase angle g. In such a configuration, most points of interest are illuminated by at least three independent sources and contours can be selected to intersect which are nearly perpendicular and where error regions are small.
这里，i1（x，y）和i2（x，y）中每10个i灰度的测量误差将表面方向限制在+6.8以内。另一方面，如果确定I1（x，y）介于0.9和1.0之间，并且确定I2（x，y）介于0.5和0.6之间，则可将表面方向确定为其真实值的25.8。这对应于第一象限中误差区域较大的梯度空间区域。这里，I1（x，y）和I2（x，y）中每10个灰度级的测量误差仅将表面方向限制在+25.80以内。
图9重复了图8的示例，但第二个光源与第一个光源的方位角相隔180度。在这种配置中，梯度空间的第二和第四象限的误差区域最小。使用两个以上光源的组合可被安排以达到所需的整体精度。一种想法是选择四个方向的照明，相对于观察者的方位均匀分布，并且具有相对较大的相位角G。在这种配置中，大多数兴趣点都由至少三个独立的光源照明，并且可以选择轮廓相交，它们几乎垂直，并且误差区域很小。

CONCLUSIONSS
Surface orientation can be determined from the image intensities obtained under a fixed imaging geometry but with varying lighting conditions. Photometric methods for determining surface orientation can be considered complementary to methods based on the identification of corresponding points in two images taken from different viewpoints:

1. Traditional stereo allows the accurate determination of distances to objects. Photometric stereo is best when the surface gradient is to be found
2. Traditional stereo works well on rough surfaces with discontinuities in surface orientation. Photometric stereo works best on smooth surfaces with few discontinuities
3. Traditional stereo works well on textured surfaces with vary surface reflectance Photometric stereo is best when applied to surfaces with uniform surface properties

Photometric stereo does have some unique advantages:1. Since the images are obtained from the same point of view there is no difficulty identifying corresponding points in the two images. This is the major computational task in traditional stereo2. Under appropriate circumstances, the surface reflectance factor can be found because the effect of surface orientation on image intensity can be removed. Traditional stereo provides no such capabilit3. Describing object shape in terms of surface orientation is preferable in a number of situations to description in terms of range or altitude above a reference plane Photometric stereo depends on a detailed understanding of the imaging process. In addition, the imaging instrument must be of high caliber so that the gray levels produced can be dependably related to scene radiance. Fortunately, our understanding of image formation and the physics of light reflection has advanced sufficiently, and the quality of imaging devices is now high enough to make this endeavor feasible
结论：
表面方向可以通过在固定的成像几何体下获得的图像强度来确定，但在不同的照明条件下。用于确定表面方向的光度法可被视为是对基于从不同视点拍摄的两幅图像中相应点的识别的方法的补充：
1。传统的立体音响可以精确地确定到物体的距离。当要找到表面梯度时，光度立体效果最好。
2。传统的立体音响适用于表面方向不连续的粗糙表面。光度学立体效果最好的平滑表面，很少有间断。
三。传统立体适用于具有不同表面反射比的纹理表面。当应用于具有均匀表面特性的表面时，立体效果最好。

光度立体确实有一些独特的优点：1。由于图像是从同一角度获得的，因此在两个图像中识别相应的点没有困难。这是传统立体声2的主要计算任务。在适当的条件下，由于可以去除表面方向对图像强度的影响，可以得到表面反射率因子。传统立体声无法提供这种功能3。在许多情况下，根据表面方向描述物体形状比根据参考平面以上的范围或高度来描述更可取，光度立体取决于对成像过程的详细理解。此外，成像仪器必须是高口径的，这样产生的灰度水平就可以可靠地与场景辐射有关。幸运的是，我们对图像形成和光反射物理的理解已经足够深入，成像设备的质量已经足够高，使这项工作成为可能。
Figure 1. Defining the three angles i, a and g. The incident angle i is the angle between the incident ray and the surface normal. The emergent angle e is the angle between the emergent ray and the surface normal. The phase angle g is the angle between the incident and emergent rays.
图1.定义三个角度i，a和g。入射角i是入射光线和表面法向之间的角度。出射角e是出射光线与表面法向的夹角。相位角G是入射光线和出射光线之间的角度。
Figure 2. Characterizing Image projections (a) illustrates the well-known perspective projection. [Note: to avoid Image inversion, it is convenient to assume that the image plane lies in front of the lens rather than behind it] For objects that are small relative to the viewing distance, the image projection can be modeled as the orthographic projection illustrated in(b). In an orthographic projection ,the focal length f is infinite so that all rays from object to image are parallel .
图2.表征图像投影（a）说明了众所周知的透视投影。[注：为了避免图像反转，可以方便地假设图像平面位于透镜的前面而不是后面]对于相对视距较小的物体，可以将图像投影建模为（b）中所示的正交投影。在正射投影中，焦距f是无穷大的，因此从物体到图像的所有光线都是平行的。