/**
* Extract a 96 byte descriptor.
*/
public void extract(GaussianScaleSpacePyramid pyramid,
DogFeaturePointStack points, FreakFeaturePointStack store)
{
ExtractFREAK84(store, pyramid, points, mPointRing0, mPointRing1,
mPointRing2, mPointRing3, mPointRing4, mPointRing5,
mSigmaCenter, mSigmaRing0, mSigmaRing1, mSigmaRing2,
mSigmaRing3, mSigmaRing4, mSigmaRing5, mExpansionFactor);
}
/**
* Detect scale-invariant feature points given a pyramid.
* @param _i_dog_feature_points
* 検出したDOG特徴点
*/
public void detect(GaussianScaleSpacePyramid i_pyramid, DogFeaturePointStack i_dog_feature_points)
{
//clean up 1st feature stack
DogFeaturePointStack tmp_fp = this._tmp_fps;
tmp_fp.clear();
// Compute Laplacian images (DoG)
this.mLaplacianPyramid.compute(i_pyramid);
// Detect minima and maximum in Laplacian images
this.extractFeatures(i_pyramid, this.mLaplacianPyramid, tmp_fp);
// Sub-pixel refinement
this.findSubpixelLocations(i_pyramid, tmp_fp);
// Compute the gradient pyramid
this.mOrientationAssignment.computeGradients(i_pyramid);
AreaBuckit abuckit = this.mBuckets;
if (tmp_fp.getLength() <= abuckit._buckit.Length)
{
//特徴点の数が要求数以下なら全てのポイントを使う。
for (int i = 0; i < tmp_fp.getLength(); i++)
{
this.addFeatureOrientations(i_pyramid, tmp_fp.getItem(i), i_dog_feature_points);
}
}
else
{
//特徴点を選別(Prune features)
// Clear the previous state
abuckit.clear();
// Insert each features into a bucket
for (int i = 0; i < tmp_fp.getLength(); i++)
{
DogFeaturePoint p = tmp_fp.getItem(i);
abuckit.put(p.x, p.y, i, Math.Abs(p.score));
}
// Compute an orientation for each feature point
for (int i = 0; i < abuckit._buckit.Length; i++)
{
if (abuckit._buckit[i].first == 0)
{
continue;
}
this.addFeatureOrientations(i_pyramid, tmp_fp.getItem(abuckit._buckit[i].second), i_dog_feature_points);
}
}
return;
}
public static NyARNftFreakFsetFile genFeatureSet3(NyARNftIsetFile i_iset_file)
{
int max_features = 500;
DogFeaturePointStack _dog_feature_points = new DogFeaturePointStack(max_features);
FreakFeaturePointStack query_keypoint = new FreakFeaturePointStack(max_features);
//
List <NyARNftFreakFsetFile.RefDataSet> refdataset = new List <NyARNftFreakFsetFile.RefDataSet>();
List <NyARNftFreakFsetFile.ImageInfo> imageinfo = new List <NyARNftFreakFsetFile.ImageInfo>();
for (int ii = 0; ii < i_iset_file.items.Length; ii++)
{
NyARNftIsetFile.ReferenceImage rimg = i_iset_file.items[ii];
FREAKExtractor mFeatureExtractor = new FREAKExtractor();
int octerves = BinomialPyramid32f.octavesFromMinimumCoarsestSize(rimg.width, rimg.height, 8);
BinomialPyramid32f _pyramid = new BinomialPyramid32f(rimg.width, rimg.height, octerves, 3);
DoGScaleInvariantDetector _dog_detector = new DoGScaleInvariantDetector(rimg.width, rimg.height, octerves, 3, 3, 4, max_features);
//RefDatasetの作成
_pyramid.build(NyARGrayscaleRaster.createInstance(rimg.width, rimg.height, NyARBufferType.INT1D_GRAY_8, rimg.img));
// Detect feature points
_dog_feature_points.clear();
_dog_detector.detect(_pyramid, _dog_feature_points);
// Extract features
query_keypoint.clear();
mFeatureExtractor.extract(_pyramid, _dog_feature_points, query_keypoint);
for (int i = 0; i < query_keypoint.getLength(); i++)
{
FreakFeaturePoint ffp = query_keypoint.getItem(i);
NyARNftFreakFsetFile.RefDataSet rds = new NyARNftFreakFsetFile.RefDataSet();
rds.pageNo = 1;
rds.refImageNo = ii;
rds.coord2D.setValue(ffp.x, ffp.y);
rds.coord3D.setValue((ffp.x + 0.5f) / rimg.dpi * 25.4f, ((rimg.height - 0.5f) - ffp.y) / rimg.dpi * 25.4f);
rds.featureVec.angle = ffp.angle;
rds.featureVec.maxima = ffp.maxima ? 1 : 0;
rds.featureVec.scale = ffp.scale;
ffp.descripter.getValueLe(rds.featureVec.v);
refdataset.Add(rds);
}
imageinfo.Add(new NyARNftFreakFsetFile.ImageInfo(rimg.width, rimg.height, ii));
}
NyARNftFreakFsetFile.PageInfo[] pi = new NyARNftFreakFsetFile.PageInfo[1];
pi[0] = new NyARNftFreakFsetFile.PageInfo(1, imageinfo.ToArray());
return(new NyARNftFreakFsetFile(refdataset.ToArray(), pi));
}
/**
* Sub-pixel refinement.
*/
private void findSubpixelLocations(GaussianScaleSpacePyramid pyramid, DogFeaturePointStack i_dog_fp)
{
int num_points;
double laplacianSqrThreshold;
double hessianThreshold;
num_points = 0;
laplacianSqrThreshold = (this.mLaplacianThreshold * this.mLaplacianThreshold);
double te = (mEdgeThreshold + 1);
hessianThreshold = ((te * te) / mEdgeThreshold);
for (int i = 0; i < i_dog_fp.getLength(); i++)
{
DogFeaturePoint kp = i_dog_fp.getItem(i);
//assert kp.scale < mLaplacianPyramid.numScalePerOctave();
// ASSERT(kp.scale < mLaplacianPyramid.numScalePerOctave(),
// "Feature point scale is out of bounds");
int lap_index = kp.octave * mLaplacianPyramid.numScalePerOctave() + kp.scale;
// Get Laplacian images
LaplacianImage lap0 = mLaplacianPyramid.get(lap_index - 1);
LaplacianImage lap1 = mLaplacianPyramid.get(lap_index);
LaplacianImage lap2 = mLaplacianPyramid.get(lap_index + 1);
// Compute the Hessian
if (!this.updateLocation(kp, lap0, lap1, lap2))
{
continue;
}
if (Math.Abs(kp.edge_score) < hessianThreshold && (kp.score * kp.score) >= laplacianSqrThreshold &&
kp.x >= 0 && kp.x < mLaplacianPyramid.get(0).getWidth() && kp.y >= 0 &&
kp.y < mLaplacianPyramid.get(0).getHeight())
{
// Update the sigma
kp.sigma = pyramid.effectiveSigma(kp.octave, kp.sp_scale);
i_dog_fp.swap(i, num_points++);
}
}
i_dog_fp.setLength(num_points);
}
/**
* Extract the descriptors for all the feature points.
*/
private void ExtractFREAK84(FreakFeaturePointStack store,
GaussianScaleSpacePyramid pyramid, DogFeaturePointStack points,
double[] points_ring0, double[] points_ring1, double[] points_ring2,
double[] points_ring3, double[] points_ring4, double[] points_ring5,
double sigma_center, double sigma_ring0, double sigma_ring1,
double sigma_ring2, double sigma_ring3, double sigma_ring4,
double sigma_ring5, double expansion_factor)
{
// ASSERT(pyramid, "Pyramid is NULL");
// ASSERT(store.size() == points.size(),
// "Feature store has not been allocated");
for (int i = 0; i < points.getLength(); i++)
{
FreakFeaturePoint sp = store.prePush();
if (sp == null)
{
prepush_wawning();
break;
}
DogFeaturePoint pt = points.getItem(i);
if (!ExtractFREAK84(sp.descripter,
pyramid, pt, points_ring0, points_ring1,
points_ring2, points_ring3, points_ring4, points_ring5,
sigma_center, sigma_ring0, sigma_ring1, sigma_ring2,
sigma_ring3, sigma_ring4, sigma_ring5, expansion_factor
))
{
store.pop();
continue;
}
sp.angle = pt.angle;
sp.x = pt.x;
sp.y = pt.y;
sp.scale = pt.sigma;
sp.maxima = pt.score > 0;
// store.point(num_points).set(points[i]);
}
}
private void addFeatureOrientations(GaussianScaleSpacePyramid i_pyramid, DogFeaturePoint dfp, DogFeaturePointStack i_ot_fps)
{
double[] tmp = this._addFeatureOrientations_tmp;
double x, y, s;
// Down sample the point to the detected octave
bilinear_downsample_point(tmp, dfp.x, dfp.y, dfp.sigma, dfp.octave);
x = tmp[0];
y = tmp[1];
s = tmp[2];
// Downsampling the point can cause (x,y) to leave the image bounds
// by
// a tiny amount. Here we just clip it to be within the image
// bounds.
x = ClipScalar(x, 0, i_pyramid.get(dfp.octave, 0).getWidth() - 1);
y = ClipScalar(y, 0, i_pyramid.get(dfp.octave, 0).getHeight() - 1);
// Compute dominant orientations
int num_angles = mOrientationAssignment.compute(dfp.octave, dfp.scale, x, y, s, this.mOrientations);
// Create a feature point for each angle
for (int j = 0; j < num_angles; j++)
{
// Copy the feature point
DogFeaturePoint fp = i_ot_fps.prePush();
if (fp == null)
{
//中断
// prepush_warning();
break;
}
fp.x = dfp.x;
fp.y = dfp.y;
fp.octave = dfp.octave;
fp.scale = dfp.scale;
fp.sp_scale = dfp.sp_scale;
fp.score = dfp.score;
fp.sigma = dfp.sigma;
fp.edge_score = dfp.edge_score;
fp.angle = mOrientations[j];
}
return;
}
/**
* Extract the minima/maxima.
*/
private void extractFeatures(GaussianScaleSpacePyramid pyramid, DoGPyramid laplacian, DogFeaturePointStack i_dog_fp)
{
double laplacianSqrThreshold = (this.mLaplacianThreshold * this.mLaplacianThreshold);
for (int i = 1; i < mLaplacianPyramid.size() - 1; i++)
{
LaplacianImage im0 = laplacian.get(i - 1);
LaplacianImage im1 = laplacian.get(i);
LaplacianImage im2 = laplacian.get(i + 1);
double[] im0b = (double[])im0.getBuffer();
double[] im1b = (double[])im1.getBuffer();
double[] im2b = (double[])im2.getBuffer();
int octave = laplacian.octaveFromIndex((int)i);
int scale = laplacian.scaleFromIndex((int)i);
if (im0.getWidth() == im1.getWidth() && im0.getWidth() == im2.getWidth())
{ // All images are the
// same size
// ASSERT(im0.height() == im1.height(), "Height is inconsistent");
// ASSERT(im0.height() == im2.height(), "Height is inconsistent");
int width_minus_1 = im1.getWidth() - 1;
int heigh_minus_1 = im1.getHeight() - 1;
for (int row = 1; row < heigh_minus_1; row++)
{
int im0_ym1 = im0.get(row - 1);
int im0_y = im0.get(row);
int im0_yp1 = im0.get(row + 1);
int im1_ym1 = im1.get(row - 1);
int im1_y = im1.get(row);
int im1_yp1 = im1.get(row + 1);
int im2_ym1 = im2.get(row - 1);
int im2_y = im2.get(row);
int im2_yp1 = im2.get(row + 1);
for (int col = 1; col < width_minus_1; col++)
{
double value = im1b[im1_y + col];
// Check laplacian score
if ((value * value) < laplacianSqrThreshold)
{
continue;
}
bool extrema = false;
if (value > im0b[im0_ym1 + col - 1] && value > im0b[im0_ym1 + col] &&
value > im0b[im0_ym1 + col + 1] && value > im0b[im0_y + col - 1] &&
value > im0b[im0_y + col] && value > im0b[im0_y + col + 1] &&
value > im0b[im0_yp1 + col - 1] && value > im0b[im0_yp1 + col] &&
value > im0b[im0_yp1 + col + 1] &&
/* im1 - 8 evaluations */
value > im1b[im1_ym1 + col - 1] && value > im1b[im1_ym1 + col] &&
value > im1b[im1_ym1 + col + 1] && value > im1b[im1_y + col - 1] &&
value > im1b[im1_y + col + 1] && value > im1b[im1_yp1 + col - 1] &&
value > im1b[im1_yp1 + col] && value > im1b[im1_yp1 + col + 1] &&
/* im2 - 9 evaluations */
value > im2b[im2_ym1 + col - 1] && value > im2b[im2_ym1 + col] &&
value > im2b[im2_ym1 + col + 1] && value > im2b[im2_y + col - 1] &&
value > im2b[im2_y + col] && value > im2b[im2_y + col + 1] &&
value > im2b[im2_yp1 + col - 1] && value > im2b[im2_yp1 + col] &&
value > im2b[im2_yp1 + col + 1])
{
extrema = true;
}
else if (value < im0b[im0_ym1 + col - 1] && value < im0b[im0_ym1 + col] &&
value < im0b[im0_ym1 + col + 1] && value < im0b[im0_y + col - 1] &&
value < im0b[im0_y + col] && value < im0b[im0_y + col + 1] &&
value < im0b[im0_yp1 + col - 1] && value < im0b[im0_yp1 + col] &&
value < im0b[im0_yp1 + col + 1] &&
/* im1 - 8 evaluations */
value < im1b[im1_ym1 + col - 1] && value < im1b[im1_ym1 + col] &&
value < im1b[im1_ym1 + col + 1] && value < im1b[im1_y + col - 1] &&
value < im1b[im1_y + col + 1] && value < im1b[im1_yp1 + col - 1] &&
value < im1b[im1_yp1 + col] && value < im1b[im1_yp1 + col + 1] &&
/* im2 - 9 evaluations */
value < im2b[im2_ym1 + col - 1] && value < im2b[im2_ym1 + col] &&
value < im2b[im2_ym1 + col + 1] && value < im2b[im2_y + col - 1] &&
value < im2b[im2_y + col] && value < im2b[im2_y + col + 1] &&
value < im2b[im2_yp1 + col - 1] && value < im2b[im2_yp1 + col] &&
value < im2b[im2_yp1 + col + 1])
{
extrema = true;
}
if (extrema)
{
DogFeaturePoint fp = i_dog_fp.prePush();
if (fp == null)
{
prepush_warning();
break;
}
fp.octave = octave;
fp.scale = scale;
fp.score = value;
fp.sigma = pyramid.effectiveSigma(octave, scale);
double[] tmp = new double[2];
bilinear_upsample_point(tmp, col, row, octave);
fp.x = tmp[0];
fp.y = tmp[1];
}
}
}
}
else if (im0.getWidth() == im1.getWidth() && (im1.getWidth() >> 1) == im2.getWidth())
{
int end_x = (int)Math.Floor(((im2.getWidth() - 1) - 0.5f) * 2.0f + 0.5f);
int end_y = (int)Math.Floor(((im2.getHeight() - 1) - 0.5f) * 2.0f + 0.5f);
for (int row = 2; row < end_y; row++)
{
int im0_ym1 = im0.get(row - 1);
int im0_y = im0.get(row);
int im0_yp1 = im0.get(row + 1);
int im1_ym1 = im1.get(row - 1);
int im1_y = im1.get(row);
int im1_yp1 = im1.get(row + 1);
for (int col = 2; col < end_x; col++)
{
double value = im1b[im1_y + col];
// Check laplacian score
if ((value * value) < laplacianSqrThreshold)
{
continue;
}
// Compute downsampled point location
double ds_x = col * 0.5f - 0.25f;
double ds_y = row * 0.5f - 0.25f;
bool extrema = false;
if (
/* im0 - 9 evaluations */
value > im0b[im0_ym1 + col - 1] && value > im0b[im0_ym1 + col] &&
value > im0b[im0_ym1 + col + 1] && value > im0b[im0_y + col - 1] &&
value > im0b[im0_y + col] && value > im0b[im0_y + col + 1] &&
value > im0b[im0_yp1 + col - 1] && value > im0b[im0_yp1 + col] &&
value > im0b[im0_yp1 + col + 1] &&
/* im1 - 8 evaluations */
value > im1b[im1_ym1 + col - 1] && value > im1b[im1_ym1 + col] &&
value > im1b[im1_ym1 + col + 1] && value > im1b[im1_y + col - 1] &&
value > im1b[im1_y + col + 1] && value > im1b[im1_yp1 + col - 1] &&
value > im1b[im1_yp1 + col] && value > im1b[im1_yp1 + col + 1] &&
/* im2 - 9 evaluations */
value > im2.bilinearInterpolation(ds_x - 0.5f, ds_y - 0.5f) &&
value > im2.bilinearInterpolation(ds_x, ds_y - 0.5f) &&
value > im2.bilinearInterpolation(ds_x + 0.5f, ds_y - 0.5f) &&
value > im2.bilinearInterpolation(ds_x - 0.5f, ds_y) &&
value > im2.bilinearInterpolation(ds_x, ds_y) &&
value > im2.bilinearInterpolation(ds_x + 0.5f, ds_y) &&
value > im2.bilinearInterpolation(ds_x - 0.5f, ds_y + 0.5f) &&
value > im2.bilinearInterpolation(ds_x, ds_y + 0.5f) &&
value > im2.bilinearInterpolation(ds_x + 0.5f, ds_y + 0.5f))
{
extrema = true;
}
else if (
/* im0 - 9 evaluations */
value < im0b[im0_ym1 + col - 1] && value < im0b[im0_ym1 + col] &&
value < im0b[im0_ym1 + col + 1] && value < im0b[im0_y + col - 1] &&
value < im0b[im0_y + col] && value < im0b[im0_y + col + 1] &&
value < im0b[im0_yp1 + col - 1] && value < im0b[im0_yp1 + col] &&
value < im0b[im0_yp1 + col + 1] &&
/* im1 - 8 evaluations */
value < im1b[im1_ym1 + col - 1] && value < im1b[im1_ym1 + col] &&
value < im1b[im1_ym1 + col + 1] && value < im1b[im1_y + col - 1] &&
value < im1b[im1_y + col + 1] && value < im1b[im1_yp1 + col - 1] &&
value < im1b[im1_yp1 + col] && value < im1b[im1_yp1 + col + 1] &&
/* im2 - 9 evaluations */
value < im2.bilinearInterpolation(ds_x - 0.5f, ds_y - 0.5f) &&
value < im2.bilinearInterpolation(ds_x, ds_y - 0.5f) &&
value < im2.bilinearInterpolation(ds_x + 0.5f, ds_y - 0.5f) &&
value < im2.bilinearInterpolation(ds_x - 0.5f, ds_y) &&
value < im2.bilinearInterpolation(ds_x, ds_y) &&
value < im2.bilinearInterpolation(ds_x + 0.5f, ds_y) &&
value < im2.bilinearInterpolation(ds_x - 0.5f, ds_y + 0.5f) &&
value < im2.bilinearInterpolation(ds_x, ds_y + 0.5f) &&
value < im2.bilinearInterpolation(ds_x + 0.5f, ds_y + 0.5f))
{
extrema = true;
}
if (extrema)
{
DogFeaturePoint fp = i_dog_fp.prePush();
if (fp == null)
{
prepush_warning();
break;
}
fp.octave = octave;
fp.scale = scale;
fp.score = value;
fp.sigma = pyramid.effectiveSigma(octave, scale);
double[] tmp = new double[2];
bilinear_upsample_point(tmp, col, row, octave);
fp.x = tmp[0];
fp.y = tmp[1];
}
}
}
}
else if ((im0.getWidth() >> 1) == im1.getWidth() && (im0.getWidth() >> 1) == im2.getWidth())
{
int width_minus_1 = im1.getWidth() - 1;
int height_minus_1 = im1.getHeight() - 1;
for (int row = 1; row < height_minus_1; row++)
{
int im1_ym1 = im1.get(row - 1);
int im1_y = im1.get(row);
int im1_yp1 = im1.get(row + 1);
int im2_ym1 = im2.get(row - 1);
int im2_y = im2.get(row);
int im2_yp1 = im2.get(row + 1);
for (int col = 1; col < width_minus_1; col++)
{
double value = im1b[im1_y + col];
// Check laplacian score
if ((value * value) < laplacianSqrThreshold)
{
continue;
}
double us_x = (col << 1) + 0.5f;
double us_y = (row << 1) + 0.5f;
bool extrema = false;
if (value > im1b[im1_ym1 + col - 1] && value > im1b[im1_ym1 + col] &&
value > im1b[im1_ym1 + col + 1] && value > im1b[im1_y + col - 1] &&
value > im1b[im1_y + col + 1] && value > im1b[im1_yp1 + col - 1] &&
value > im1b[im1_yp1 + col] && value > im1b[im1_yp1 + col + 1] &&
/* im2 - 9 evaluations */
value > im2b[im2_ym1 + col - 1] && value > im2b[im2_ym1 + col] &&
value > im2b[im2_ym1 + col + 1] && value > im2b[im2_y + col - 1] &&
value > im2b[im2_y + col] && value > im2b[im2_y + col + 1] &&
value > im2b[im2_yp1 + col - 1] && value > im2b[im2_yp1 + col] &&
value > im2b[im2_yp1 + col + 1] &&
/* im2 - 9 evaluations */
value > im0.bilinearInterpolation(us_x - 2.0f, us_y - 2.0f) &&
value > im0.bilinearInterpolation(us_x, us_y - 2.0f) &&
value > im0.bilinearInterpolation(us_x + 2.0f, us_y - 2.0f) &&
value > im0.bilinearInterpolation(us_x - 2.0f, us_y) &&
value > im0.bilinearInterpolation(us_x, us_y) &&
value > im0.bilinearInterpolation(us_x + 2.0f, us_y) &&
value > im0.bilinearInterpolation(us_x - 2.0f, us_y + 2.0f) &&
value > im0.bilinearInterpolation(us_x, us_y + 2.0f) &&
value > im0.bilinearInterpolation(us_x + 2.0f, us_y + 2.0f))
{
extrema = true;
}
else if (value < im1b[im1_ym1 + col - 1] && value < im1b[im1_ym1 + col] &&
value < im1b[im1_ym1 + col + 1] && value < im1b[im1_y + col - 1] &&
value < im1b[im1_y + col + 1] && value < im1b[im1_yp1 + col - 1] &&
value < im1b[im1_yp1 + col] && value < im1b[im1_yp1 + col + 1] &&
/* im2 - 9 evaluations */
value < im2b[im2_ym1 + col - 1] && value < im2b[im2_ym1 + col] &&
value < im2b[im2_ym1 + col + 1] && value < im2b[im2_y + col - 1] &&
value < im2b[im2_y + col] && value < im2b[im2_y + col + 1] &&
value < im2b[im2_yp1 + col - 1] && value < im2b[im2_yp1 + col] &&
value < im2b[im2_yp1 + col + 1] &&
/* im2 - 9 evaluations */
value < im0.bilinearInterpolation(us_x - 2.0f, us_y - 2.0f) &&
value < im0.bilinearInterpolation(us_x, us_y - 2.0f) &&
value < im0.bilinearInterpolation(us_x + 2.0f, us_y - 2.0f) &&
value < im0.bilinearInterpolation(us_x - 2.0f, us_y) &&
value < im0.bilinearInterpolation(us_x, us_y) &&
value < im0.bilinearInterpolation(us_x + 2.0f, us_y) &&
value < im0.bilinearInterpolation(us_x - 2.0f, us_y + 2.0f) &&
value < im0.bilinearInterpolation(us_x, us_y + 2.0f) &&
value < im0.bilinearInterpolation(us_x + 2.0f, us_y + 2.0f))
{
extrema = true;
}
if (extrema)
{
DogFeaturePoint fp = i_dog_fp.prePush();
if (fp == null)
{
prepush_warning();
break;
}
fp.octave = octave;
fp.scale = scale;
fp.score = value;
fp.sigma = pyramid.effectiveSigma(octave, scale);
double[] tmp = new double[2];
bilinear_upsample_point(tmp, col, row, octave);
fp.x = tmp[0];
fp.y = tmp[1];
}
}
}
}
}
return;
}
