Improving accuracy for cumulative matrices (Normalization)
To avoid degeneration of transform matrix select one axis as main. I usually chose Z as it is usually view or forward direction in my apps. Then exploit cross product to recompute/normalize the rest of axises (which should be perpendicular to each other and unless scale is used then also unit size). This can be done only for orthonormal matrices so no skew or projections ... Orthogonal matrices must be scaled to orthonormal then inverted and then scaled back to make this usable.
You do not need to do this after every operation just make a counter of operations done on each matrix and if some threshold crossed then normalize it and reset counter.
To detect degeneration of such matrices you can test for orthogonality by dot product between any two axises (should be zero or very near it). For orthonormal matrices you can test also for unit size of axis direction vectors ...
Here is how my transform matrix normalization looks like (for orthonormal matrices) in C++:
double reper::rep[16];
void reper::orto(int test)
{
double x[3],y[3],z[3];
if ((cnt>=_reper_max_cnt)||(test))
{
axisx_get(x);
axisy_get(y);
axisz_get(z);
vector_one(z,z);
vector_mul(x,y,z);
vector_one(x,x);
vector_mul(y,z,x);
vector_one(y,y);
axisx_set(x);
axisy_set(y);
axisz_set(z);
cnt=0;
}
}
void reper::axisx_get(double *p)
{
p[0]=rep[0];
p[1]=rep[1];
p[2]=rep[2];
}
void reper::axisx_set(double *p)
{
rep[0]=p[0];
rep[1]=p[1];
rep[2]=p[2];
cnt=_reper_max_cnt;
}
void reper::axisy_get(double *p)
{
p[0]=rep[4];
p[1]=rep[5];
p[2]=rep[6];
}
void reper::axisy_set(double *p)
{
rep[4]=p[0];
rep[5]=p[1];
rep[6]=p[2];
cnt=_reper_max_cnt;
}
void reper::axisz_get(double *p)
{
p[0]=rep[ 8];
p[1]=rep[ 9];
p[2]=rep[10];
}
void reper::axisz_set(double *p)
{
rep[ 8]=p[0];
rep[ 9]=p[1];
rep[10]=p[2];
cnt=_reper_max_cnt;
}
The vector operations looks like this:
void vector_one(double *c,double *a)
{
double l=divide(1.0,sqrt((a[0]*a[0])+(a[1]*a[1])+(a[2]*a[2])));
c[0]=a[0]*l;
c[1]=a[1]*l;
c[2]=a[2]*l;
}
void vector_mul(double *c,double *a,double *b)
{
double q[3];
q[0]=(a[1]*b[2])-(a[2]*b[1]);
q[1]=(a[2]*b[0])-(a[0]*b[2]);
q[2]=(a[0]*b[1])-(a[1]*b[0]);
for(int i=0;i<3;i++) c[i]=q[i];
}
Improving accuracy for non cumulative matrices
Your only choice is use at least double accuracy of your matrices. Safest is to use GLM or your own matrix math based at least on double data type (like my reper class).
Cheap alternative is using double precision functions like
glTranslated
glRotated
glScaled
...
which in some cases helps but is not safe as OpenGL implementation can truncate it to float. Also there are no 64 bit HW interpolators yet so all iterated results between pipeline stages are truncated to floats.
Sometimes relative reference frame helps (so keep operations on similar magnitude values) for example see:
ray and ellipsoid intersection accuracy improvement
Also In case you are using own matrix math functions you have to consider also the order of operations so you always lose smallest amount of accuracy possible.
Pseudo inverse matrix
In some cases you can avoid computing of inverse matrix by determinants or Horner scheme or Gauss elimination method because in some cases you can exploit the fact that Transpose of orthonormal rotational matrix is also its inverse. Here is how it is done:
void matrix_inv(GLfloat *a,GLfloat *b)
{
GLfloat x,y,z;
a[ 0]=b[ 0];
a[ 5]=b[ 5];
a[10]=b[10];
x=b[1]; a[1]=b[4]; a[4]=x;
x=b[2]; a[2]=b[8]; a[8]=x;
x=b[6]; a[6]=b[9]; a[9]=x;
a[ 3]=b[ 3];
a[ 7]=b[ 7];
a[11]=b[11];
a[15]=b[15];
x=(a[ 0]*b[12])+(a[ 4]*b[13])+(a[ 8]*b[14]);
y=(a[ 1]*b[12])+(a[ 5]*b[13])+(a[ 9]*b[14]);
z=(a[ 2]*b[12])+(a[ 6]*b[13])+(a[10]*b[14]);
a[12]=-x;
a[13]=-y;
a[14]=-z;
}
So rotational part of the matrix is transposed, projection stays as was and origin position is recomputed so A*inverse(A)=unit_matrix This function is written so it can be used as in-place so calling
GLfloat a[16]={values,...}
matrix_inv(a,a);
lead to valid results too. This way of computing Inverse is quicker and numerically safer as it pends much less operations (no recursion or reductions no divisions). Of coarse this works only for orthonormal homogenuous 4x4 matrices !!!*
Detection of wrong inverse
So if you got matrix A and its inverse B then:
A*B = C = ~unit_matrix
So multiply both matrices and check for unit matrix...
- abs sum of all non diagonal elements of
C should be close to 0.0
- all diagonal elements of
C should be close to +1.0
near = distance(camera, centerOfWorld) - radusOfWorld和far = distance(camera, centerOfWorld) + radusOfWorld。当相机在包围盒内部时，near=nearMin（例如=1个单位，以查看细节），而far=2*radiusOfWorld。在这种情况下，我不会关注Z-fighting。 - Ripi2