void LMS::fit( //fit sample
float &out_m, //output line
float &out_c) {
inT32 index; //of median
inT32 trials; //no of medians
float test_m, test_c; //candidate line
float test_error; //error of test line
switch (samplecount) {
case 0:
m = 0.0f; //no info
c = 0.0f;
line_error = 0.0f;
break;
case 1:
m = 0.0f;
c = samples[0].y (); //horiz thru pt
line_error = 0.0f;
break;
case 2:
if (samples[0].x () != samples[1].x ()) {
m = (samples[1].y () - samples[0].y ())
/ (samples[1].x () - samples[0].x ());
c = samples[0].y () - m * samples[0].x ();
}
else {
m = 0.0f;
c = (samples[0].y () + samples[1].y ()) / 2;
}
line_error = 0.0f;
break;
default:
pick_line(m, c); //use pts at random
compute_errors(m, c); //from given line
index = choose_nth_item (samplecount / 2, errors, samplecount);
line_error = errors[index];
for (trials = 1; trials < lms_line_trials; trials++) {
//random again
pick_line(test_m, test_c);
compute_errors(test_m, test_c);
index = choose_nth_item (samplecount / 2, errors, samplecount);
test_error = errors[index];
if (test_error < line_error) {
//find least median
line_error = test_error;
m = test_m;
c = test_c;
}
}
}
fitted = TRUE;
out_m = m;
out_c = c;
a = 0;
}
void LMS::constrained_fit( //fit sample
float fixed_m, //forced gradient
float &out_c) {
inT32 index; //of median
inT32 trials; //no of medians
float test_c; //candidate line
static uinT16 seeds[3] = { SEED1, SEED2, SEED3 };
//for nrand
float test_error; //error of test line
m = fixed_m;
switch (samplecount) {
case 0:
c = 0.0f;
line_error = 0.0f;
break;
case 1:
//horiz thru pt
c = samples[0].y () - m * samples[0].x ();
line_error = 0.0f;
break;
case 2:
c = (samples[0].y () + samples[1].y ()
- m * (samples[0].x () + samples[1].x ())) / 2;
line_error = m * samples[0].x () + c - samples[0].y ();
line_error *= line_error;
break;
default:
index = (inT32) nrand48 (seeds) % samplecount;
//compute line
c = samples[index].y () - m * samples[index].x ();
compute_errors(m, c); //from given line
index = choose_nth_item (samplecount / 2, errors, samplecount);
line_error = errors[index];
for (trials = 1; trials < lms_line_trials; trials++) {
index = (inT32) nrand48 (seeds) % samplecount;
test_c = samples[index].y () - m * samples[index].x ();
//compute line
compute_errors(m, test_c);
index = choose_nth_item (samplecount / 2, errors, samplecount);
test_error = errors[index];
if (test_error < line_error) {
//find least median
line_error = test_error;
c = test_c;
}
}
}
fitted = TRUE;
out_c = c;
a = 0;
}
// Compute all the cross product distances of the points from the line
// and return the true squared upper quartile distance.
double DetLineFit::ComputeErrors(const ICOORD start, const ICOORD end,
int* distances) {
ICOORDELT_IT it(&pt_list_);
ICOORD line_vector = end;
line_vector -= start;
// Compute the distance of each point from the line.
int pt_index = 0;
for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
ICOORD pt_vector = *it.data();
pt_vector -= start;
// Compute |line_vector||pt_vector|sin(angle between)
int dist = line_vector * pt_vector;
if (dist < 0)
dist = -dist;
distances[pt_index++] = dist;
}
// Now get the upper quartile distance.
int index = choose_nth_item(3 * pt_index / 4, distances, pt_index,
sizeof(distances[0]), CompareInts);
double dist = distances[index];
// The true distance is the square root of the dist squared / the
// squared length of line_vector (which is the dot product with itself)
// Don't bother with the square root. Just return the square distance.
return dist * dist / (line_vector % line_vector);
}
float LMS::compute_quadratic_errors( //fit sample
float outlier_threshold, //min outlier
double line_a,
float line_m, //input gradient
float line_c) {
inT32 outlier_count; //total outliers
inT32 index; //picked point
inT32 error_count; //no in total
double total_error; //summed squares
total_error = 0;
outlier_count = 0;
error_count = 0;
for (index = 0; index < samplecount; index++) {
errors[error_count] = line_c + samples[index].x ()
* (line_m + samples[index].x () * line_a) - samples[index].y ();
errors[error_count] *= errors[error_count];
if (errors[error_count] > outlier_threshold) {
outlier_count++;
errors[samplecount - outlier_count] = errors[error_count];
}
else {
total_error += errors[error_count++];
}
}
if (outlier_count * 3 < error_count)
return total_error / error_count;
else {
index = choose_nth_item (outlier_count / 2,
errors + samplecount - outlier_count,
outlier_count);
//median outlier
return errors[samplecount - outlier_count + index];
}
}
DLLSYM inT32
choose_nth_item ( //fast median
inT32 index, //index to choose
void *array, //array of items
inT32 count, //no of items
size_t size, //element size
//comparator
int (*compar) (const void *, const void *)
) {
int result; //of compar
inT32 next_sample; //next one to do
inT32 next_lesser; //space for new
inT32 prev_greater; //last one saved
inT32 equal_count; //no of equal ones
inT32 pivot; //proposed median
if (count <= 1)
return 0;
if (count == 2) {
if (compar (array, (char *) array + size) < 0) {
return index >= 1 ? 1 : 0;
}
else {
return index >= 1 ? 0 : 1;
}
}
if (index < 0)
index = 0; //ensure lergal
else if (index >= count)
index = count - 1;
pivot = (inT32) (rand () % count);
swap_entries (array, size, pivot, 0);
next_lesser = 0;
prev_greater = count;
equal_count = 1;
for (next_sample = 1; next_sample < prev_greater;) {
result =
compar ((char *) array + size * next_sample,
(char *) array + size * next_lesser);
if (result < 0) {
swap_entries (array, size, next_lesser++, next_sample++);
//shuffle
}
else if (result > 0) {
prev_greater--;
swap_entries(array, size, prev_greater, next_sample);
}
else {
equal_count++;
next_sample++;
}
}
if (index < next_lesser)
return choose_nth_item (index, array, next_lesser, size, compar);
else if (index < prev_greater)
return next_lesser; //in equal bracket
else
return choose_nth_item (index - prev_greater,
(char *) array + size * prev_greater,
count - prev_greater, size,
compar) + prev_greater;
}
DLLSYM inT32 choose_nth_item( //fast median
inT32 index, //index to choose
float *array, //array of items
inT32 count //no of items
) {
inT32 next_sample; //next one to do
inT32 next_lesser; //space for new
inT32 prev_greater; //last one saved
inT32 equal_count; //no of equal ones
float pivot; //proposed median
float sample; //current sample
if (count <= 1)
return 0;
if (count == 2) {
if (array[0] < array[1]) {
return index >= 1 ? 1 : 0;
}
else {
return index >= 1 ? 0 : 1;
}
}
else {
if (index < 0)
index = 0; //ensure lergal
else if (index >= count)
index = count - 1;
equal_count = (inT32) (rand() % count);
pivot = array[equal_count];
//fill gap
array[equal_count] = array[0];
next_lesser = 0;
prev_greater = count;
equal_count = 1;
for (next_sample = 1; next_sample < prev_greater;) {
sample = array[next_sample];
if (sample < pivot) {
//shuffle
array[next_lesser++] = sample;
next_sample++;
}
else if (sample > pivot) {
prev_greater--;
//juggle
array[next_sample] = array[prev_greater];
array[prev_greater] = sample;
}
else {
equal_count++;
next_sample++;
}
}
for (next_sample = next_lesser; next_sample < prev_greater;)
array[next_sample++] = pivot;
if (index < next_lesser)
return choose_nth_item (index, array, next_lesser);
else if (index < prev_greater)
return next_lesser; //in equal bracket
else
return choose_nth_item (index - prev_greater,
array + prev_greater,
count - prev_greater) + prev_greater;
}
}
