Examples with DMatrixRMaj org.ejml.data.DMatrixRMaj used on opensource projects

Example 76 with DMatrixRMaj

use of org.ejml.data.DMatrixRMaj in project BoofCV by lessthanoptimal.

the class TestProjectiveReconstructionByFactorization method perfect_input_badDepths.

/**
 * All data is perfect, but 1 is used for the depth estimate
 */
@Test
void perfect_input_badDepths() {
    int numViews = 8;
    int numFeatures = 10;
    simulate(numViews, numFeatures, false);
    ProjectiveStructureByFactorization alg = new ProjectiveStructureByFactorization();
    alg.initialize(features3D.size(), projections.size());
    // depth isn't known so just set it to 1. it could easily converge to a poor local optimal
    alg.setAllDepths(1);
    for (int viewIdx = 0; viewIdx < projections.size(); viewIdx++) {
        alg.setPixels(viewIdx, observations.get(viewIdx));
    }
    assertTrue(alg.process());
    DMatrixRMaj P = new DMatrixRMaj(3, 4);
    Point4D_F64 X = new Point4D_F64();
    int total = 0;
    for (int viewIdx = 0; viewIdx < numViews; viewIdx++) {
        alg.getCameraMatrix(viewIdx, P);
        for (int featureIdx = 0; featureIdx < numFeatures; featureIdx++) {
            alg.getFeature3D(featureIdx, X);
            Point2D_F64 expected = observations.get(viewIdx).get(featureIdx);
            Point3D_F64 xh = PerspectiveOps.renderPixel(P, X, (Point3D_F64) null);
            Point2D_F64 found = new Point2D_F64(xh.x / xh.z, xh.y / xh.z);
            // System.out.println(expected+" "+found);
            if (expected.distance(found) <= 2)
                total++;
        }
    }
    // see if a large number of solutions are within 2 pixels
    assertTrue(total >= 0.95 * numViews * numFeatures);
}

Example 77 with DMatrixRMaj

use of org.ejml.data.DMatrixRMaj in project BoofCV by lessthanoptimal.

the class TestProjectiveStructureFromHomographies method constructLinearSystem.

/**
 * Tests linear system by seeing if zero is returned when truth is passed in. Input camera matrix is
 * K*[R,T] to make the match much easier.
 */
@Test
void constructLinearSystem() {
    int numViews = 4;
    int numFeatures = 7;
    simulate(numViews, numFeatures, true);
    ProjectiveStructureFromHomographies alg = new ProjectiveStructureFromHomographies();
    List<DMatrixRMaj> homographies = new ArrayList<>();
    for (int i = 0; i < projections.size(); i++) {
        DMatrixRMaj R = new DMatrixRMaj(3, 3);
        CommonOps_DDRM.extract(projections.get(i), 0, 0, R);
        homographies.add(R);
    }
    List<List<PointIndex2D_F64>> observations = convertToPointIndex();
    alg.computeConstants(homographies, observations, numFeatures);
    alg.constructLinearSystem(homographies, observations);
    DMatrixRMaj X = new DMatrixRMaj(alg.numUnknown, 1);
    for (int i = 0; i < features3D.size(); i++) {
        int row = i * 3;
        X.data[row] = features3D.get(i).x;
        X.data[row + 1] = features3D.get(i).y;
        X.data[row + 2] = features3D.get(i).z;
    }
    for (int i = 0; i < homographies.size(); i++) {
        // homography is from view 0 to view i
        DMatrixRMaj P = projections.get(i);
        int row = features3D.size() * 3 + i * 3;
        X.data[row] = P.get(0, 3);
        X.data[row + 1] = P.get(1, 3);
        X.data[row + 2] = P.get(2, 3);
    }
    DMatrixRMaj B = new DMatrixRMaj(alg.numEquations, 1);
    CommonOps_DDRM.mult(alg.A, X, B);
    assertEquals(0, NormOps_DDRM.normF(B), UtilEjml.TEST_F64);
}

Example 78 with DMatrixRMaj

use of org.ejml.data.DMatrixRMaj in project BoofCV by lessthanoptimal.

the class TestProjectiveStructureFromHomographies method perfect_input.

/**
 * Given perfect input see if it can create camera matrices which will return the original observations
 */
@Test
void perfect_input() {
    int numViews = 8;
    int numFeatures = 20;
    simulate(numViews, numFeatures, true);
    ProjectiveStructureFromHomographies alg = new ProjectiveStructureFromHomographies();
    assertTrue(alg.proccess(computeHomographies(), convertToPointIndex(), numFeatures));
    DMatrixRMaj P = new DMatrixRMaj(3, 4);
    Point3D_F64 X = new Point3D_F64();
    for (int viewIdx = 0; viewIdx < numViews; viewIdx++) {
        alg.getCameraMatrix(viewIdx, P);
        for (int featureIdx = 0; featureIdx < numFeatures; featureIdx++) {
            alg.getFeature3D(featureIdx, X);
            Point2D_F64 expected = observations.get(viewIdx).get(featureIdx);
            Point2D_F64 x = PerspectiveOps.renderPixel(P, X, (Point2D_F64) null);
            assertTrue(expected.distance(x) < UtilEjml.TEST_F64_SQ);
        }
    }
}

Example 79 with DMatrixRMaj

use of org.ejml.data.DMatrixRMaj in project BoofCV by lessthanoptimal.

the class TestRefineTwoViewPinholeRotation method evaluateImperfect.

private void evaluateImperfect(CameraPinhole intrinsic1, CameraPinhole intrinsic2, List<AssociatedPair> pairs) {
    var alg = new RefineTwoViewPinholeRotation();
    // Apply some constraints. Yes this test isn't exhaustive at all.
    alg.assumeUnityAspect = true;
    alg.zeroSkew = true;
    alg.assumeSameIntrinsics = false;
    alg.knownFocalLength = false;
    // Focal length and rotation are going to be way off as a test
    CameraPinhole found1 = new CameraPinhole(200, 200, 0.0, 530, 530, 1000, 1000);
    CameraPinhole found2 = new CameraPinhole(1000, 1000, 0.0, 430, 430, 800, 800);
    DMatrixRMaj R = CommonOps_DDRM.identity(3);
    alg.converge.maxIterations = 40;
    // alg.setVerbose(System.out, null);
    assertTrue(alg.refine(pairs, R, found1, found2));
    // Should be significantly better
    assertTrue(alg.errorAfter * 1e5 <= alg.errorBefore);
    // Check constraints
    assertEquals(found1.fx, found1.fy);
    assertEquals(found2.fx, found2.fy);
    assertEquals(0.0, found1.skew);
    assertEquals(0.0, found2.skew);
    // Based on manual inspection there seems to be redundancy in these equations some place that I've
    // not yet dug out. The error is zero but K1 and K2 are significantly different. R is "close"
    // very crude test for focal length
    assertEquals(intrinsic1.fx, found1.fx, 150);
    assertEquals(intrinsic2.fx, found2.fy, 150);
    // Compute the number of radians different the two rotations are
    DMatrixRMaj diffR = CommonOps_DDRM.multTransA(view1_to_view2.R, R, null);
    double angle = ConvertRotation3D_F64.matrixToRodrigues(diffR, null).theta;
    assertEquals(0.0, angle, 0.05);
}

Example 80 with DMatrixRMaj

use of org.ejml.data.DMatrixRMaj in project BoofCV by lessthanoptimal.

the class TestRefineTwoViewPinholeRotation method perfectInput.

/**
 * Give it perfect input and see if it screws things up
 */
@Test
void perfectInput() {
    CameraPinhole intrinsic1 = new CameraPinhole(400, 410, 0.1, 500, 550, 1000, 1000);
    CameraPinhole intrinsic2 = new CameraPinhole(600, 550, 0.02, 400, 440, 800, 800);
    List<AssociatedPair> pairs = renderPairs(intrinsic1, intrinsic2, 50, 0.0);
    var alg = new RefineTwoViewPinholeRotation();
    // turn off all assumptions
    alg.assumeUnityAspect = false;
    alg.zeroSkew = false;
    alg.assumeSameIntrinsics = false;
    CameraPinhole found1 = new CameraPinhole(intrinsic1);
    CameraPinhole found2 = new CameraPinhole(intrinsic2);
    DMatrixRMaj R = view1_to_view2.R.copy();
    assertTrue(alg.refine(pairs, R, found1, found2));
    // Score shouldn't get worse, but since the input is perfect it might not get better
    assertTrue(alg.errorAfter <= alg.errorBefore);
    // should be very very similar
    assertTrue(found1.isEquals(intrinsic1, 1e-6));
    assertTrue(found2.isEquals(intrinsic2, 1e-6));
    assertTrue(MatrixFeatures_DDRM.isIdentical(view1_to_view2.R, R, 1e-6));
}