CGII/framework/include/cgv/math/plane.h
2018-05-17 16:01:02 +02:00

199 lines
3.8 KiB
C++

#pragma once
#include <cgv/math/vec.h>
#include <cgv/math/point_operations.h>
#include <cgv/math/eig.h>
#include <algorithm>
namespace cgv{
namespace math{
/**
* A plane is defined as a vector (a,b,c,d) => a*x1 + b*x2 +c*x3 +d = 0
*/
///evaluate implicit plane equation at x =(x1,x2,x3)
///return value should be zero on plane
template <typename T>
T plane_val(const vec<T>& plane,const vec<T>& x)
{
assert(plane.size()-1 == x.size());
T val=0;
for(unsigned i = 0;i< x.size(); i++)
val += plane(i)*x(i);
return val+plane(plane.size()-1);
}
///evaluate plame equation on multiple positions xs
template <typename T>
vec<T> plane_val(const vec<T>& plane,const mat<T>& xs)
{
assert(plane.size()-1 == xs.nrows());
vec<T> vals(xs.ncols());
for(unsigned i = 0;i< xs.ncols(); i++)
vals(i) = plane_val(plane,xs.col(i));
return vals;
}
//construct implicit plane from 3 points
template <typename T>
vec<T> plane_fit(const vec<T>& p1,const vec<T>& p2,const vec<T>& p3)
{
vec<T> plane(4);
double l_nml;
vec<T> nml = cross(normalize(p2-p1),normalize(p3-p1));
l_nml = length(nml);
assert(l_nml != 0);
nml/=l_nml;
plane.set(nml(0),nml(1),nml(2),-dot(p1,nml));
return plane;
}
//fit implicit plane into multiple points (total least squares)
template <typename T>
vec<T> tls_plane_fit(const mat<T>& points)
{
vec<T> plane(4);
mat<T> covmat,v;
diag_mat<T> d;
vec<T> mean;
covmat_and_mean(points,covmat, mean);
eig_sym(covmat,v,d);
// T l_nml;
vec<T> nml = normalize(v.col(2));
plane.set(nml(0),nml(1),nml(2),-dot(mean,nml));
return plane;
}
///ransac plane fit
///p_out... outlier prob
///d_max... threshold distance
///p_surety... surety to compute number needed samples
///if m_sac flag is true m-estimator cost function is used
///if loransac flag is true
template <typename T>
vec<T> ransac_plane_fit(const mat<T>& points,const T p_out=0.8, const T d_max=0.001, const T p_surety = 0.99, bool msac=true)
{
assert(points.nrows == 3);
vec<T> plane(4);
vec<unsigned int> ind;
vec<T> p1,p2,p3,dists(points.ncols());
cgv::math::random rg;
unsigned n = points.ncols;
unsigned max_iter = num_ransac_iterations(3,p_out,psurety);
unsigned num_inlier = (unsigned)::ceil((1-p_out)*points.ncols());
T error = std::numeric_limits<T>::max();
for(unsigned i = 0; i < max_iter; i++)
{
//random sample
do
{
rg.uniform_nchoosek(n,3,ind);
p1 = points.col(ind(0));
p2 = points.col(ind(1));
p3 = points.col(ind(2));
}
while(fabs(length(cross(q2-q1,q3-q1)))<0.001);
//fit
vec<T> pl = plane_fit(p1,p2,p3);
//consensus
int n_inlier=0;
T e = 0;
for(unsigned i = 0; i < points.ncols(); i++)
{
dists(i) = std::abs(plane_val(pl,points.col(i)));
if(dists(i) < d_max)
{
n_inlier++;
if(msac)
e += dists(i);
}
else
{
e += dmax;
}
}
//local reoptimize
if(loransac && n_inlier > 3)
{
mat<T> inliers(3,n_inlier);
int j=0;
n_inlier=0;
for(unsigned i = 0; i < points.ncols(); i++)
{
if(dists(i) < d_max)
{
n_inlier++;
inliers.set_col(j,points.col(i));
j++;
}
}
pl = plane_fit(inliers)
//recompute error
for(unsigned i = 0; i < points.ncols(); i++)
{
dists(i) = std::abs(plane_val(pl,points.col(i)));
if(dists(i) < d_max)
{
if(msac)
e += dists(i);
}
else
{
e += dmax;
}
}
}
//remember best plane
if(error > e)
{
error = e;
plane = pl;
}
if(n_inlier >= num_inlier)
{
p_out = (T)1-(T)n_inlier/(T)points.ncols();
max_iter = num_ransac_iterations(3,p_out,psurety);
}
}
return plane;
}
}
}