PBRT_V2 总结记录 <36> IrregIsotropicBRDF

概述 

IrregIsotropicBRDF 的作用：

（利用测量的数据来模拟 表面的反射，例如：其实就是在类中保存了BRDF的很多采样点，根据wi,wo 来对 获得采样点之间的插值之后的数据，这些数据就是对应的BRDF值）

Using measured data about the reflection properties of real surfaces is one of the most
effective approaches for rendering realistic materials. One way to do so is to use the
measured data to set the parameter values for a parameterized BRDF like the Torrance–
Sparrow model.

However, parameterized BRDF models are often not flexible enough to model the full
complexity of scattering characteristics of many interesting surfaces. If accurate measured
reflection data is available for a surface of interest, then directly using it for rendering
can faithfully(如实地) re-create the surface’s appearance. In this section, we’ll implement
BxDFs that interpolate sampled BRDF values directly. While this is a memory-intensive(内存密集型)
approach for scenes that use many different measured BRDFs, it works well for reflection
distributions that don’t fit analytic models well and can provide a reference for comparison
to other approaches.

Rendering with measured BRDF data does have challenges. It can be difficult to make
adjustments for artistic purposes.If there is excessive
error in the data, results may be poor as well: especially for highly specular and
anisotropic surfaces, the 4D BRDF must be densely sampled, leading to a large amount
of data to store and many measurements to be taken. Finally, BRDFs may have considerable
variation in their values, which adds to the challenges of acquiring accurate BRDF
data sets.

Measured BRDF data may come in one of two forms: as regularly spaced tabularized(列成表格)
data, or as a large number of irregularly spaced individual samples. Regularly spaced
data makes efficient lookups possible. (For example, if we have a 4D table of values for
an anisotropic BRDF, indexed by (θi , φi , θo, φo), then, given a pair of directions, we just
need to index into the table and interpolate between the values.) However, it can be difficult
to acquire BRDF values with such a regular spacing. Acquiring good measurements
for directions close to the horizon can be particularly difficult, for example. Irregular
sample spacings avoid these issues, though are not as easy to use for rendering. (Irregular
sets of BRDF samples are often resampled in a preprocess to generate a set of regular
samples.)

Therefore, pbrt has implementations of two BxDFs for using measured reflection data for
rendering. The first, IrregIsotropicBRDF interpolates from irregularly spaced samples of
isotropic BRDFs. The second, RegularHalfangleBRDF, supports the regular sampling used
byMatusik et al.’s data set (2003a, 2003b). Both of these are used by the MeasuredMaterial
defined in Section 9.2.4.

 

 

IrregIsotropicBRDF 类

class IrregIsotropicBRDF : public BxDF {
public:
    // IrregIsotropicBRDF Public Methods
    IrregIsotropicBRDF(const KdTree<IrregIsotropicBRDFSample> *d)
        : BxDF(BxDFType(BSDF_REFLECTION | BSDF_GLOSSY)), isoBRDFData(d) { }
    Spectrum f(const Vector &wo, const Vector &wi) const;
private:
    // IrregIsotropicBRDF Private Data
    const KdTree<IrregIsotropicBRDFSample> *isoBRDFData;
};

类的作用：

（IrregIsotropicBRDF  中存储 不整齐的 各向同性 的BRDF的采样点，给一对(wi,wo)的话，就会根据这对方向去采样点数组中找到最适合的BRDF采样值，利用这些采样值进行插件得到最终的BRDF值 ）

IrregIsotropicBRDF supports irregularly sampled isotropic BRDF data. Given a pair of
directions, it finds a few BRDF samples that are nearby in the space of incident and
outgoing directions and interpolates between their values, based on how close they are
to the actual directions.

 

1. 构造函数

IrregIsotropicBRDF(const KdTree<IrregIsotropicBRDFSample> *d)
        : BxDF(BxDFType(BSDF_REFLECTION | BSDF_GLOSSY)), isoBRDFData(d) { }

const KdTree<IrregIsotropicBRDFSample> *isoBRDFData;

细节：

a.

（用 面的公式来表示一个BRDF的采样点）

Point BRDFRemap(const Vector &wo, const Vector &wi);
struct IrregIsotropicBRDFSample {
    IrregIsotropicBRDFSample(const Point &pp, const Spectrum &vv)
        : p(pp), v(vv) { }
    IrregIsotropicBRDFSample() { }
    Point p;
    Spectrum v;
};

Point BRDFRemap(const Vector &wo, const Vector &wi) {
    float cosi = CosTheta(wi), coso = CosTheta(wo);
    float sini = SinTheta(wi), sino = SinTheta(wo);
    float phii = SphericalPhi(wi), phio = SphericalPhi(wo);
    float dphi = phii - phio;
    if (dphi < 0.) dphi += 2.f * M_PI;
    if (dphi > 2.f * M_PI) dphi -= 2.f * M_PI;
    if (dphi > M_PI) dphi = 2.f * M_PI - dphi;
    return Point(sini * sino, dphi / M_PI, cosi * coso);
}

We’d therefore like to use a mapping function for isotropic BRDFs that takes a 4D pair of
directions ωi and ωo and converts them into a point in a three-dimensional space such
that the distance between two points is meaningful, the isotropy of the BRDF is reflected,
and reciprocity is represented. For the IrregIsotropicBRDF, we’ll implement a mapping
proposed by Marschner, denoted by φ3 in his thesis(Marschner 1998; Section 5.6.3):

where  Δφ = (φi − φo), and Δφ is remapped to be in the range [0, π]. (Thanks to the
isotropy assumption, Δφ values in [π, 2π] are equivalent to 2π − Δφ, which is in
[0, π].)

 

This mapping accounts for isotropy and reciprocity, and the distance between two three
tuples in the mapping’s range has a meaningful relationship to the distance between their
respective pairs of directions. The BRDFRemap() function implements this mapping.

 

b.

（用 kd-tree 来保存BRDF的采样值）

When a IrregIsotropicBRDF is created by the MeasuredMaterial::GetBSDF() method, it
is given a pointer to a three-dimensional kd-tree that stores the BRDF samples, where
the sample positions are represented with the values returned by BRDFRemap(). Note that
this kd-tree is managed by the MeasuredMaterial and is only used for lookups (i.e., in a
read-only manner) by the code here.

 

2. 

Spectrum f(const Vector &wo, const Vector &wi) const;

作用：

（利用wo，wi，来重构一个 point出来，就利用 point 去 kd-tree 来查找 多个 采样点 ，再进行加权平均，求出平均值作为 BRDF的值）

The IrregIsotropicBRDF::f() method first remaps the given pair of directions and then
performs a series of kd-tree lookups, searching with a progressively wider search radius
for nearby points until a few samples have been found. In general, it’s preferable to
smoothly interpolate between a few BRDF samples rather than just find the single closest
sample for the lookup here.

 

细节：

Spectrum IrregIsotropicBRDF::f(const Vector &wo,
                               const Vector &wi) const {
    Point m = BRDFRemap(wo, wi);
    float lastMaxDist2 = .001f;
    while (true) {
        // Try to find enough BRDF samples around _m_ within search radius
        IrregIsoProc proc;
        float maxDist2 = lastMaxDist2;
        isoBRDFData->Lookup(m, proc, maxDist2);
        if (proc.nFound > 2 || lastMaxDist2 > 1.5f)
            return proc.v.Clamp() / proc.sumWeights;
        lastMaxDist2 *= 2.f;
    }
}

struct IrregIsoProc {
    // IrregIsoProc Public Methods
    IrregIsoProc() { sumWeights = 0.f; nFound = 0; }
    void operator()(const Point &p, const IrregIsotropicBRDFSample &sample,
                    float d2, float &maxDist2) {
        float weight = expf(-100.f * d2);
        v += weight * sample.v;
        sumWeights += weight;
        ++nFound;
    }
    Spectrum v;
    float sumWeights;
    int nFound;
};

a.

IrregIsoProc is the callback structure used by the KdTree::Lookup() method for each
sample within the search radius. (The squared search radius is provided to the Lookup()
method in maxDist2.) If not enough points are found, the squared search distance is
doubled, and another lookup is performed.

IrregIsoProc processes each BRDF sample within the search radius, accumulating a
weighted sum of samples close to the lookup point.

BRDF sample values are weighted with an ad hoc exponential(指数) falloff based on the squared
distance between the lookup point and the sample point. The sum of these weights is
accumulated so that the final value can be normalized.