@Test
public void elementMultConjB() {
InterleavedF64 a = new InterleavedF64(width, height, 2);
InterleavedF64 b = new InterleavedF64(width, height, 2);
InterleavedF64 c = new InterleavedF64(width, height, 2);
ImageMiscOps.fillUniform(a, rand, -10, 10);
ImageMiscOps.fillUniform(b, rand, -10, 10);
ImageMiscOps.fillUniform(c, rand, -10, 10);
CirculantTracker.elementMultConjB(a, b, c);
for (int y = 0; y < height; y++) {
for (int x = 0; x < width; x++) {
Complex64F aa = new Complex64F(a.getBand(x, y, 0), a.getBand(x, y, 1));
Complex64F bb = new Complex64F(b.getBand(x, y, 0), b.getBand(x, y, 1));
Complex64F cc = new Complex64F();
ComplexMath64F.conj(bb, bb);
ComplexMath64F.mult(aa, bb, cc);
double foundReal = c.getBand(x, y, 0);
double foundImg = c.getBand(x, y, 1);
assertEquals(cc.real, foundReal, 1e-4);
assertEquals(cc.imaginary, foundImg, 1e-4);
}
}
}
@Test
public void computeAlphas() {
InterleavedF64 yf = new InterleavedF64(width, height, 2);
InterleavedF64 kf = new InterleavedF64(width, height, 2);
InterleavedF64 alphaf = new InterleavedF64(width, height, 2);
ImageMiscOps.fillUniform(yf, rand, -10, 10);
ImageMiscOps.fillUniform(kf, rand, -10, 10);
ImageMiscOps.fillUniform(alphaf, rand, -10, 10);
float lambda = 0.01f;
CirculantTracker.computeAlphas(yf, kf, lambda, alphaf);
for (int y = 0; y < height; y++) {
for (int x = 0; x < width; x++) {
Complex64F a = new Complex64F(yf.getBand(x, y, 0), yf.getBand(x, y, 1));
Complex64F b = new Complex64F(kf.getBand(x, y, 0) + lambda, kf.getBand(x, y, 1));
Complex64F c = new Complex64F();
ComplexMath64F.div(a, b, c);
double foundReal = alphaf.getBand(x, y, 0);
double foundImg = alphaf.getBand(x, y, 1);
assertEquals(c.real, foundReal, 1e-4);
assertEquals(c.imaginary, foundImg, 1e-4);
}
}
}
private void setTargetLocation(int x, int y) {
ImageMiscOps.fillUniform(image, rand, 0, 1);
ImageMiscOps.fillRectangle(image, 100, cornerX, cornerY, 20, 20);
pyramid.process(image);
for (int i = 0; i < pyramid.getNumLayers(); i++) {
GradientSobel.process(
pyramid.getLayer(i),
derivX[i],
derivY[i],
new ImageBorder1D_F32(BorderIndex1D_Extend.class));
}
}
public BenchmarkConvolveNormalizeEdge() {
imgInt8 = new ImageUInt8(imgWidth, imgHeight);
imgInt16 = new ImageSInt16(imgWidth, imgHeight);
out_I32 = new ImageSInt32(imgWidth, imgHeight);
out_I16 = new ImageSInt16(imgWidth, imgHeight);
out_I8 = new ImageUInt8(imgWidth, imgHeight);
imgFloat32 = new ImageFloat32(imgWidth, imgHeight);
out_F32 = new ImageFloat32(imgWidth, imgHeight);
Random rand = new Random(234234);
ImageMiscOps.fillUniform(imgInt8, rand, 0, 100);
ImageMiscOps.fillUniform(imgInt16, rand, 0, 200);
ImageMiscOps.fillUniform(imgFloat32, rand, 0, 200);
}
protected void checkMotion(double tranX, double tranY, double rot) {
ImageUInt8 frame0 = new ImageUInt8(320, 240);
ImageUInt8 frame1 = new ImageUInt8(320, 240);
ImageMiscOps.fillUniform(frame0, rand, 0, 256);
double c = Math.cos(rot);
double s = Math.sin(rot);
DistortImageOps.affine(frame0, frame1, TypeInterpolate.BILINEAR, c, -s, s, c, tranX, tranY);
SfotConfig config = new SfotConfig();
ImageGradient<ImageUInt8, ImageSInt16> gradient =
FactoryDerivative.sobel(ImageUInt8.class, ImageSInt16.class);
SparseFlowObjectTracker<ImageUInt8, ImageSInt16> alg =
new SparseFlowObjectTracker<ImageUInt8, ImageSInt16>(
config, ImageUInt8.class, ImageSInt16.class, gradient);
RectangleRotate_F64 region0 = new RectangleRotate_F64(120, 140, 30, 40, 0.1);
RectangleRotate_F64 region1 = new RectangleRotate_F64();
alg.init(frame0, region0);
assertTrue(alg.update(frame1, region1));
double expectedX = c * region0.cx - s * region0.cy + tranX;
double expectedY = s * region0.cx + c * region0.cy + tranY;
double expectedYaw = UtilAngle.bound(region0.theta + rot);
assertEquals(expectedX, region1.cx, 0.5);
assertEquals(expectedY, region1.cy, 0.5);
assertEquals(expectedYaw, region1.theta, 0.01);
}
@Test
public void performLearning() {
float interp_factor = 0.075f;
ImageFloat32 a = new ImageFloat32(20, 25);
ImageFloat32 b = new ImageFloat32(20, 25);
ImageMiscOps.fill(a, 100);
ImageMiscOps.fill(b, 200);
CirculantTracker<ImageFloat32> alg =
new CirculantTracker<ImageFloat32>(1f / 16, 0.2, 1e-2, 0.075, 1.0, 64, 255, interp);
alg.initialize(a, 0, 0, 20, 25);
// copy its internal value
ImageFloat64 templateC = new ImageFloat64(alg.template.width, alg.template.height);
templateC.setTo(alg.template);
// give it two images
alg.performLearning(b);
// make sure the images aren't full of zero
assertTrue(Math.abs(ImageStatistics.sum(templateC)) > 0.1);
assertTrue(Math.abs(ImageStatistics.sum(alg.template)) > 0.1);
int numNotSame = 0;
// the result should be an average of the two
for (int i = 0; i < a.data.length; i++) {
if (Math.abs(a.data[i] - alg.templateNew.data[i]) > 1e-4) numNotSame++;
// should be more like the original one than the new one
double expected =
templateC.data[i] * (1 - interp_factor) + interp_factor * alg.templateNew.data[i];
double found = alg.template.data[i];
assertEquals(expected, found, 1e-4);
}
// make sure it is actually different
assertTrue(numNotSame > 100);
}
/**
* Creates a random image and looks for corners in it. Sees if the naive and fast algorithm
* produce exactly the same results.
*/
@Test
public void compareToNaive() {
GrayU8 img = new GrayU8(width, height);
ImageMiscOps.fillUniform(img, new Random(0xfeed), 0, 100);
GrayS16 derivX = new GrayS16(img.getWidth(), img.getHeight());
GrayS16 derivY = new GrayS16(img.getWidth(), img.getHeight());
GradientSobel.process(img, derivX, derivY, new ImageBorder1D_S32(BorderIndex1D_Extend.class));
BoofTesting.checkSubImage(this, "compareToNaive", true, derivX, derivY);
}
@Override
public void renderTarget(ImageFloat32 original, List<CalibrationObservation> solutions) {
ImageMiscOps.fill(original, 255);
int numRows = config.numRows * 2 - 1;
int numCols = config.numCols * 2 - 1;
int square = original.getWidth() / (Math.max(numRows, numCols) + 4);
int targetWidth = square * numCols;
int targetHeight = square * numRows;
int x0 = (original.width - targetWidth) / 2;
int y0 = (original.height - targetHeight) / 2;
for (int i = 0; i < numRows; i += 2) {
int y = y0 + i * square;
for (int j = 0; j < numCols; j += 2) {
int x = x0 + j * square;
ImageMiscOps.fillRectangle(original, 0, x, y, square, square);
}
}
int pointsRow = numRows + 1;
int pointsCol = numCols + 1;
CalibrationObservation set = new CalibrationObservation();
int gridIndex = 0;
for (int i = 0; i < pointsRow; i++) {
for (int j = 0; j < pointsCol; j++, gridIndex++) {
double y = y0 + i * square;
double x = x0 + j * square;
set.add(new Point2D_F64(x, y), gridIndex);
}
}
solutions.add(set);
}
@Test
public void imageDotProduct() {
ImageFloat64 a = new ImageFloat64(width, height);
ImageMiscOps.fillUniform(a, rand, 0, 10);
double total = 0;
for (int y = 0; y < height; y++) {
for (int x = 0; x < width; x++) {
total += a.get(x, y) * a.get(x, y);
}
}
double found = CirculantTracker.imageDotProduct(a);
assertEquals(total, found, 1e-8);
}
protected void performThresholding(float threshLow, float threshHigh, ImageUInt8 output) {
if (hysteresisPts != null) {
hysteresisPts.process(suppressed, direction, threshLow, threshHigh);
// if there is an output image write the contour to it
if (output != null) {
ImageMiscOps.fill(output, 0);
for (EdgeContour e : hysteresisPts.getContours()) {
for (EdgeSegment s : e.segments)
for (Point2D_I32 p : s.points) output.unsafe_set(p.x, p.y, 1);
}
}
} else {
hysteresisMark.process(suppressed, direction, threshLow, threshHigh, output);
}
}
@Test
public void compareToConvolve_I8() throws NoSuchMethodException {
CompareDerivativeToConvolution validator = new CompareDerivativeToConvolution();
validator.setTarget(
HessianThree_Standard.class.getMethod(
"process", ImageUInt8.class, ImageSInt16.class, ImageSInt16.class, ImageSInt16.class));
validator.setKernel(0, HessianThree.kernelXXYY_I32, true);
validator.setKernel(1, HessianThree.kernelXXYY_I32, false);
validator.setKernel(2, HessianThree.kernelCross_I32);
ImageUInt8 input = new ImageUInt8(width, height);
ImageMiscOps.fillUniform(input, rand, 0, 10);
ImageSInt16 derivXX = new ImageSInt16(width, height);
ImageSInt16 derivYY = new ImageSInt16(width, height);
ImageSInt16 derivXY = new ImageSInt16(width, height);
validator.compare(false, input, derivXX, derivYY, derivXY);
}
@Test
public void compareToConvolve_F32() throws NoSuchMethodException {
CompareDerivativeToConvolution validator = new CompareDerivativeToConvolution();
validator.setTarget(
GradientPrewitt.class.getMethod(
"process",
ImageFloat32.class,
ImageFloat32.class,
ImageFloat32.class,
ImageBorder_F32.class));
validator.setKernel(0, GradientPrewitt.kernelDerivX_F32);
validator.setKernel(1, GradientPrewitt.kernelDerivY_F32);
ImageFloat32 input = new ImageFloat32(width, height);
ImageMiscOps.fillUniform(input, rand, 0, 10);
ImageFloat32 derivX = new ImageFloat32(width, height);
ImageFloat32 derivY = new ImageFloat32(width, height);
validator.compare(input, derivX, derivY);
}
@Test
public void checkRender() {
// Easier to make up a plane in this direction
Se3_F64 cameraToPlane = new Se3_F64();
ConvertRotation3D_F64.eulerToMatrix(
EulerType.XYZ, UtilAngle.degreeToRadian(0), 0, 0, cameraToPlane.getR());
cameraToPlane.getT().set(0, -5, 0);
Se3_F64 planeToCamera = cameraToPlane.invert(null);
CreateSyntheticOverheadViewMS<ImageFloat32> alg =
new CreateSyntheticOverheadViewMS<ImageFloat32>(
TypeInterpolate.BILINEAR, 3, ImageFloat32.class);
alg.configure(param, planeToCamera, centerX, centerY, cellSize, overheadW, overheadH);
MultiSpectral<ImageFloat32> input =
new MultiSpectral<ImageFloat32>(ImageFloat32.class, width, height, 3);
for (int i = 0; i < 3; i++) ImageMiscOps.fill(input.getBand(i), 10 + i);
MultiSpectral<ImageFloat32> output =
new MultiSpectral<ImageFloat32>(ImageFloat32.class, overheadW, overheadH, 3);
alg.process(input, output);
for (int i = 0; i < 3; i++) {
ImageFloat32 o = output.getBand(i);
// check parts that shouldn't be in view
assertEquals(0, o.get(0, 300), 1e-8);
assertEquals(0, o.get(5, 0), 1e-8);
assertEquals(0, o.get(5, 599), 1e-8);
// check areas that should be in view
assertEquals(10 + i, o.get(499, 300), 1e-8);
}
}
@Test
public void basic() {
ImageUInt8 image = new ImageUInt8(400, 500);
int value = 200;
ImageMiscOps.fillRectangle(image, value, 20, 30, 40, 40);
PolygonEdgeScore<ImageUInt8> alg =
new PolygonEdgeScore<ImageUInt8>(2, 2, 10, value * 0.9, ImageUInt8.class);
Polygon2D_F64 polygon = new Polygon2D_F64(4);
UtilPolygons2D_F64.convert(new Rectangle2D_F64(20, 30, 60, 70), polygon);
alg.setImage(image);
assertTrue(alg.validate(polygon));
assertEquals(value, alg.getAverageEdgeIntensity(), 1e-8);
UtilPolygons2D_F64.convert(new Rectangle2D_F64(24, 30, 60, 70), polygon);
// test a negative case
assertFalse(alg.validate(polygon));
assertEquals(value * 3.0 / 4.0, alg.getAverageEdgeIntensity(), 1e-8);
}
