Mirrors/overte
3
0
Fork 0
mirror of https://github.com/overte-org/overte.git synced 2025-04-20 03:44:02 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge pull request #3510 from ey6es/metavoxels

Browse source 
Use simpler and more reliable Jacobi eigenvalue algorithm to compute SVD.
This commit is contained in:
AndrewMeadows 2014-09-30 09:00:18 -07:00
commit 418bf1ec3c
1 changed files with 28 additions and 100 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

128
interface/src/MetavoxelSystem.cpp
View file

@ -1882,114 +1882,42 @@ int VoxelAugmentVisitor::visit(MetavoxelInfo& info) {
glm::mat3 atrans = glm::transpose(a);
glm::mat3 ata = atrans * a;
// compute the SVD of ata; first, find the eigenvalues
// (see http://en.wikipedia.org/wiki/Eigenvalue_algorithm#3.C3.973_matrices)
glm::vec3 eigenvalues;
float p1 = ata[0][1] * ata[0][1] + ata[0][2] * ata[0][2] + ata[1][2] * ata[1][2];
if (p1 < EPSILON) {
eigenvalues = glm::vec3(ata[0][0], ata[1][1], ata[2][2]);
if (eigenvalues[2] < eigenvalues[1]) {
qSwap(eigenvalues[2], eigenvalues[1]);
// find the eigenvalues and eigenvectors of ata
// (see http://en.wikipedia.org/wiki/Jacobi_eigenvalue_algorithm)
glm::mat3 d = ata;
glm::quat combinedRotation;
const int MAX_ITERATIONS = 20;
for (int i = 0; i < MAX_ITERATIONS; i++) {
glm::vec3 offDiagonals = glm::abs(glm::vec3(d[1][0], d[2][0], d[2][1]));
int largestIndex = (offDiagonals[0] > offDiagonals[1]) ? (offDiagonals[0] > offDiagonals[2] ? 0 : 2) :
(offDiagonals[1] > offDiagonals[2] ? 1 : 2);
const float DESIRED_PRECISION = 0.00001f;
if (offDiagonals[largestIndex] < DESIRED_PRECISION) {
break;
}
if (eigenvalues[1] < eigenvalues[0]) {
qSwap(eigenvalues[1], eigenvalues[0]);
int largestJ = (largestIndex == 2) ? 1 : 0;
int largestI = (largestIndex == 0) ? 1 : 2;
float sjj = d[largestJ][largestJ];
float sii = d[largestI][largestI];
float angle = (sii == sjj ? PI_OVER_TWO : glm::atan(2.0f * d[largestJ][largestI], sjj - sii)) / 2.0f;
glm::quat rotation = glm::angleAxis(angle, largestIndex == 0 ? glm::vec3(0.0f, 0.0f, -1.0f) :
(largestIndex == 1 ? glm::vec3(0.0f, 1.0f, 0.0f) : glm::vec3(-1.0f, 0.0f, 0.0f)));
if (rotation.w == 0.0f) {
break;
}
if (eigenvalues[2] < eigenvalues[1]) {
qSwap(eigenvalues[2], eigenvalues[1]);
}
} else {
float q = (ata[0][0] + ata[1][1] + ata[2][2]) / 3.0f;
float d1 = ata[0][0] - q, d2 = ata[1][1] - q, d3 = ata[2][2] - q;
float p2 = d1 * d1 + d2 * d2 + d3 * d3 + 2.0f * p1;
float p = glm::sqrt(p2 / 6.0f);
glm::mat3 b = (ata - glm::mat3(q)) / p;
float r = glm::determinant(b) / 2.0f;
float phi;
if (r <= -1.0f) {
phi = PI / 3.0f;
} else if (r >= 1.0f) {
phi = 0.0f;
} else {
phi = glm::acos(r) / 3.0f;
}
eigenvalues[2] = q + 2.0f * p * glm::cos(phi);
eigenvalues[0] = q + 2.0f * p * glm::cos(phi + (2.0f * PI / 3.0f));
eigenvalues[1] = 3.0f * q - eigenvalues[0] - eigenvalues[2];
combinedRotation = glm::normalize(rotation * combinedRotation);
glm::mat3 matrix = glm::mat3_cast(combinedRotation);
d = matrix * ata * glm::transpose(matrix);
}
// form the singular matrix from the eigenvalues
glm::mat3 d;
const float MIN_SINGULAR_THRESHOLD = 0.1f;
d[0][0] = (eigenvalues[0] < MIN_SINGULAR_THRESHOLD) ? 0.0f : 1.0f / eigenvalues[0];
d[1][1] = (eigenvalues[1] < MIN_SINGULAR_THRESHOLD) ? 0.0f : 1.0f / eigenvalues[1];
d[2][2] = (eigenvalues[2] < MIN_SINGULAR_THRESHOLD) ? 0.0f : 1.0f / eigenvalues[2];
glm::mat3 m[] = { ata - glm::mat3(eigenvalues[0]), ata - glm::mat3(eigenvalues[1]),
ata - glm::mat3(eigenvalues[2]) };
// form the orthogonal matrix from the eigenvectors
// see http://www.geometrictools.com/Documentation/EigenSymmetric3x3.pdf
bool same01 = glm::abs(eigenvalues[0] - eigenvalues[1]) < EPSILON;
bool same12 = glm::abs(eigenvalues[1] - eigenvalues[2]) < EPSILON;
glm::mat3 u;
if (!(same01 && same12)) {
if (same01 || same12) {
int i = same01 ? 2 : 0;
for (int j = 0; j < 3; j++) {
glm::vec3 first = glm::vec3(m[i][0][j], m[i][1][j], m[i][2][j]);
int j2 = (j + 1) % 3;
glm::vec3 second = glm::vec3(m[i][0][j2], m[i][1][j2], m[i][2][j2]);
glm::vec3 cross = glm::cross(first, second);
float length = glm::length(cross);
if (length > EPSILON) {
u[0][i] = cross[0] / length;
u[1][i] = cross[1] / length;
u[2][i] = cross[2] / length;
break;
}
}
i = (i + 1) % 3;
for (int j = 0; j < 3; j++) {
glm::vec3 first = glm::vec3(m[i][0][j], m[i][1][j], m[i][2][j]);
float length = glm::length(first);
if (length > EPSILON) {
glm::vec3 second = glm::cross(first, glm::vec3(1.0f, 0.0f, 0.0f));
length = glm::length(second);
if (length < EPSILON) {
second = glm::cross(first, glm::vec3(0.0f, 1.0f, 0.0f));
length = glm::length(second);
}
u[0][i] = second[0] / length;
u[1][i] = second[1] / length;
u[2][i] = second[2] / length;
second = glm::normalize(glm::cross(second, first));
i = (i + 1) % 3;
u[0][i] = second[0];
u[1][i] = second[1];
u[2][i] = second[2];
break;
}
}
} else {
for (int i = 0; i < 3; i++) {
for (int j = 0; j < 3; j++) {
glm::vec3 first = glm::vec3(m[i][0][j], m[i][1][j], m[i][2][j]);
int j2 = (j + 1) % 3;
glm::vec3 second = glm::vec3(m[i][0][j2], m[i][1][j2], m[i][2][j2]);
glm::vec3 cross = glm::cross(first, second);
float length = glm::length(cross);
if (length > EPSILON) {
u[0][i] = cross[0] / length;
u[1][i] = cross[1] / length;
u[2][i] = cross[2] / length;
break;
}
}
}
}
}
d[0][0] = (d[0][0] < MIN_SINGULAR_THRESHOLD) ? 0.0f : 1.0f / d[0][0];
d[1][1] = (d[1][1] < MIN_SINGULAR_THRESHOLD) ? 0.0f : 1.0f / d[1][1];
d[2][2] = (d[2][2] < MIN_SINGULAR_THRESHOLD) ? 0.0f : 1.0f / d[2][2];
// compute the pseudo-inverse, ataplus, and use to find the minimizing solution
glm::mat3 u = glm::mat3_cast(combinedRotation);
glm::mat3 ataplus = glm::transpose(u) * d * u;
glm::vec3 solution = (ataplus * atrans * b) + center;