ProjectiveGeometry23 0.1.0__py3-none-any.whl

ProjectiveGeometry23/__init__.py

@@ -0,0 +1 @@
+# This file makes ProjectiveGeometry23 a Python package.

ProjectiveGeometry23/central_projection.py

@@ -0,0 +1,249 @@
+"""
+Linear projection from real projective three-space to real projective two-space.
+Representations of a pinhole camera as Projection matrix optionally with additional information of image size and pixelation.
+
+Author: André Aichert
+Date: June 22, 2023
+"""
+
+import numpy as np
+import scipy
+
+
+# recommended:
+# from rich import print
+# np.set_printoptions(suppress=True)
+
+class ProjectionMatrix:
+    def __init__(self, P, image_size=(400, 300), pixel_spacing=1.0):
+        """Pinhole projection model as 3x4 projection matrix with optional
+        information of detector size and pixel spacing [mm per px]."""
+        self.P = np.array(P).reshape((3, 4))
+        self.image_size = np.array(image_size)
+        self.pixel_spacing = pixel_spacing
+
+
+    @classmethod
+    def perspective_look_at(cls, eye, center=np.array([0, 0, 0]), image_size=(400, 300), fovy_rad=0.5, pixel_spacing=1.0):
+        """Create a ProjectionMatrix that looks at a center point from and eye point given field-of-view and image size.
+
+        Extrinsic parameters as rotation matix R and translation vector t:
+        - eye: Position [X,Y,Z] of the camera (camera center)
+        - center: Point the camera is looking at (target position)
+        
+        Intrinsic parameters of a pinhole camera as a 3x3 matrix K.
+        - fovy_rad: Field of view angle in radians (vertical)
+        - image_size: Width and height of the image sensor
+
+        Returns ProjectionMatrix P=K[R t] with given image_size
+        
+        TODO: optional pixel_spacing should probably affect intrinsics
+        """
+        # Up vector defining the camera's orientation
+        up=np.array([0, 1, 0])
+        
+        # Intrinsics
+        w, h = image_size
+        tanfov2 = 2 * np.tan(0.5 * fovy_rad)
+        fx, fy = h / tanfov2, h / tanfov2
+        cx, cy = 0.5 * w, 0.5 * h
+
+        K = np.array([[fx, 0, cx],
+                      [0, fy, cy],
+                      [0, 0, 1]])
+
+        # Extrinsics
+        fwd = (center - eye) / np.linalg.norm(center - eye)
+        left = np.cross(up, fwd)
+        left /= np.linalg.norm(left)
+        new_up = np.cross(fwd, left)
+
+        R = np.array([left, new_up, -fwd])
+        t = -R.dot(eye)
+        
+        # Compose projection matrix
+        P = np.zeros((3, 4))
+        P[:, :3] = K @ R
+        P[:, 3] = K @ t
+        
+        return cls(P, image_size, pixel_spacing)
+
+    def to_dict(self):
+        return {
+            "P": self.P.tolist(),
+            "image_size": self.image_size.to_dict(),
+            "pixel_spacing": self.pixel_spacing
+        }
+
+    def to_ompl(self, with_geometry=True):
+        def f2str(f):
+            return np.format_float_positional(f, trim='-')
+        s = "; ".join([" ".join([f2str(x) for x in row]) for row in self.P])
+        if with_geometry:
+            return f'#> spacing="{self.pixel_spacing}" detector_size_px="{self.image_size}"\n[{s}]'
+        else:
+            return f'[{s}]'
+        
+    @classmethod
+    def from_dict(cls, data):
+        return cls(
+            data["P"],
+            np.array(data["image_size"]),
+            data["pixel_spacing"]
+        )
+
+    def to_json(self):
+        return json.dumps(self.to_dict(), indent=2)
+
+    @classmethod
+    def from_json(cls, json_str):
+        data = json.loads(json_str)
+        return cls.from_dict(data)
+
+    def __repr__(self):
+        return f"ProjectionMatrix:\n{self.P}\nImage Size: {self.image_size}\nPixel Spacing: {self.pixel_spacing}"
+
+    def principal_ray(self):
+        return self.P[2][0:3]
+
+    def normalize(self):
+        norm_m3 = np.linalg.norm(self.principal_ray())
+        detM = np.linalg.det(self.P[0:3, 0:3])
+        if detM < 0:
+            norm_m3 *= -1;
+        self.P /= norm_m3
+
+    def getCenterOfProjection(self):
+        _, _, V = np.linalg.svd(self.P)
+        C = V[-1, :4]
+        if C[3] < -1e-12 or C[3] > 1e-12:
+            C = C / C[3]  # By definition: Camera centers are always positive points.
+        return C.reshape(-1, 1)
+
+    def getPrincipalPoint(self):
+        pp = self.P[:3, :3] @ self.P[2, :3].T
+        return pp / pp[2]
+
+    def getPrincipalRay(self):
+        m3 = self.P[2, :3]
+        if np.linalg.norm(m3) > 1e-12:
+            m3 = m3 / np.linalg.norm(m3)
+        return m3.reshape(-1, 1)
+
+    def getFocalLengthPx(self):
+        """Compute the focal length in pixels (diagonal entries of K in P=K[R t] )."""
+        self.normalize()
+        
+        m1 = self.P[0, :3]
+        m2 = self.P[1, :3]
+        m3 = self.P[2, :3]
+
+        U = np.cross(m3, m2) / np.linalg.norm(np.cross(m3, m2))
+        V = np.cross(m3, m1) / np.linalg.norm(np.cross(m3, m1))
+
+        focal_length_u_px = np.dot(m1, np.cross(V, m3))
+        # focal_length_v_px = np.dot(m2, np.cross(U, m3))  # should be identical
+
+        return focal_length_u_px
+
+    def getDetectorAxisDirections(self):
+        """Compute the two three-points where the image u- and v-axes meet infinity. Scaled to world coordinates."""
+        self.normalize()
+        
+        m1 = self.P[0, :3]
+        m2 = self.P[1, :3]
+        m3 = self.P[2, :3]
+
+        U = np.append(np.cross(m3, m2) / np.linalg.norm(np.cross(m3, m2)), 0)
+        V = np.append(np.cross(m3, m1) / np.linalg.norm(np.cross(m3, m1)), 0)
+
+        return U, V
+
+    def getDetectorAxisDirectionsPx(self):
+        """Compute the two three-points where the image u- and v-axes
+        meet infinity. Scaled to pixels."""
+        U, V = getDetectorAxisDirections()
+        return U * self.pixel_spacing, V * self.pixel_spacing
+
+    def getDetectorPlane(self, imageVPointsUp=False):
+        """ Decomposes the projection matrix to compute the equation of the
+        image/detector plane. Assumes rectangular pixels. This is identical to
+        the principal place shifted by the focal length. For left-handed
+        coordinate systems, set imageVPointsUp to True.
+        Returns the image detector plane in Hessian normal form and a boolean
+        indicating whether imageVPointsUp appears to be set correctly. """
+        # TODO: alternative (that won't know when handedness is off)
+        # focal_length_u_mm = self.getFocalLength() * self.pixel_spacing
+        K, R, t, appears_flipped = self.decomposition(imageVPointsUp)
+        focal_length_u_mm = K[0, 0] * self.pixel_spacing
+        principal_plane = self.P[2, :4].copy()
+        principal_plane /= np.linalg.norm(principal_plane[:3])
+        principal_plane[3] -= focal_length_u_mm
+        return principal_plane, appears_flipped
+
+    def decomposition(self, imageVPointsUp = False):
+        """Decompose Projection Matrix into K[R|t] using RQ-decomposition.
+        Returns K, R, t and a boolean indicating if R is left-handed.
+        For right-handed world coordinates implies imageVPointsUp is wrong."""
+        # Compute RQ decomposition of leftmost 3x3 sub-matrix of P
+        R, Q = scipy.linalg.rq(self.P[:3, :3].copy())
+        K = R
+        R = Q
+        # make diagonal of K positive
+        S = np.eye(3)
+        if K[0, 0] < 0:
+            S[0, 0] = -1
+        if K[1, 1] < 0:
+            S[1, 1] = -1
+        if imageVPointsUp:
+            S[1, 1] *= -1
+        if K[2, 2] < 0:
+            S[2, 2] = -1
+        K = K @ S
+        R = S @ R
+        # Force zero elements in K
+        K[1, 0] = K[2, 0] = K[2, 1] = 0
+        # Scale
+        K *= 1.0 / K[2, 2]
+        t = np.linalg.solve(K, self.P[:3, 3])
+        # EXPLANATION of appears_flipped:
+        # In oriented projective geometry, for a visible point x=P*X the
+        # homogeneous coordinate will be positive. A negative homogeneous
+        # coordinate implies "behind the camera". You can thus change the
+        # direction in which a camera is facing by a multiplication of the
+        # projection matrix with -1.
+        # This is relevant for the decomposition, because in practice, the
+        # coordinate system can be left-handed. In this case, the image
+        # "rotation" matrix has negative determinant. This may actually be the
+        # case, e.g. for the way pixels are stored in a BMP file, some window
+        # coordinates or Siemens 'Leonardo' style raw images. In most cases,
+        # though, that means you got the oriantation wrong.
+        # You can check using sourceDetectorGeometry. Only if your pixel
+        # spacing and imageVPointsUp values are correct, will the image plane
+        # and detector origin be on the opposite side of the object w.r.t.
+        # source position (as it should be... duh).
+        appears_flipped = np.linalg.det(R) > 0
+        return K, R, t, appears_flipped
+
+    def pseudoinverse(self):
+        return np.linalg.pinv(self.P)
+
+    def backproject(self, x):
+        """Compute Plücker Coordinates of a backprojection for a 2D image point x in homogeneous coordinates.
+        if x is a list, make sure to convert o a column vector np.array(x).reshape(-1, 1)."""
+        Pinv = self.pseudoinverse()
+        return Pinv @ x
+
+    def computeFundamentalMatrix(self, P1):
+        """Compute fundamental matrix from two projection matrices. Pseudoinverse-based implementation"""
+        if isinstance(P1, ProjectionMatrix):
+            P1 = P1.P
+        C0 = self.getCenterOfProjection()
+        e1 = P1 @ C0
+        e1 = e1.flatten() / np.linalg.norm(e1)
+        e1x = np.array([[     0, +e1[2], -e1[1]],
+                        [-e1[2],      0, +e1[0]],
+                        [+e1[1], -e1[0],      0]])
+        P0plus = self.pseudoinverse()
+
+        return  np.dot(np.dot(e1x, P1), P0plus)

ProjectiveGeometry23/estimation.py

@@ -0,0 +1,71 @@
+"""
+Direct Linear Transformation (DLT)
+Author: André Aichert
+Date: June 22, 2023
+"""
+
+
+import numpy as np
+
+def dlt_normalization(p2d, p3d, matches):
+    """Normalization of 2D and 3D point clouds. Input assumed to have homogeneous coordinate 1.
+    Purpose: De-mean and scale to +/- sqtr(2), sqrt(3) respectively. """
+    # Compute mean
+    mean2d = np.array([0, 0])
+    mean3d = np.array([0, 0, 0])
+    for i, j in matches:
+        mean2d += p2d[i][:2]
+        mean3d += p3d[j][:3]
+    mean2d /= len(match)
+    mean3d /= len(match)
+
+    # Compute size
+    s2d = 0
+    s3d = 0
+    for i, j in matches:
+        s2d += np.linalg.norm(p2d[i][:2] - mean2d)
+        s3d += np.linalg.norm(p3d[j][:3] - mean3d)
+    s2d *= np.sqrt(2) / len(match)
+    s3d *= np.sqrt(3) / len(match)
+
+    # Compose normalization matrices
+    normalization_2d = np.matmul(scale(s2d, s2d), translation(-mean2d[0], -mean2d[1]))
+    normalization_3d = np.matmul(scale(s3d, s3d, s3d), translation(-mean3d[0], -mean3d[1], -mean3d[2]))
+
+    return normalization_2d, normalization_3d
+
+
+def dlt(x, X, match=None):
+    """Direct Linear Transformation (DLT) for projection matrices."""
+    # Optional match parameter
+    if not match and len(X) == len(x):
+        match = {i: i for i in range(len(X))}
+
+    # Check for insufficient data
+    if len(match) < 6:
+        return np.zeros((3, 4))
+
+    # Normalization of input data
+    N_px, N_mm = dlt_normalization(x, X, match)
+
+    # Build homogeneous system matrix from point matches
+    A = np.zeros((2 * len(match), 12))
+    for k, (key, value) in enumerate(match.items()):
+        # Normalize input points
+        x_norm = np.dot(N_px, x[key])
+        X_norm = np.dot(N_mm, X[value])
+        # Write two rows in A (we get two independent equations from one point match)
+        A[2 * k, 4:8] = x_norm[2] * X_norm
+        A[2 * k + 1, 0:4] = -x_norm[2] * X_norm
+        A[2 * k, 8:12] = -x_norm[1] * X_norm
+        A[2 * k + 1, 8:12] = x_norm[0] * X_norm
+
+    # Solve and reshape
+    _, _, V = np.linalg.svd(A)
+    p = V[-1] / V[-1, -1]
+    P_norm = p.reshape((3, 4))
+
+    # Denormalize
+    P = np.dot(np.dot(np.linalg.inv(N_px), P_norm), N_mm)
+    P /= np.linalg.norm(P[:, :3], axis=0)
+    return P

ProjectiveGeometry23/homography.py

@@ -0,0 +1,136 @@
+"""
+Useful homographies on real projective two- and three-space.
+Author: André Aichert
+Date: June 22, 2023
+"""
+
+import numpy as np
+
+
+def translation2d(t):
+    """Homogeneous 2D translation."""
+    T = np.array([
+        [1, 0, t[0]],
+        [0, 1, t[1]],
+        [0, 0,    1]
+    ])
+    return T
+
+
+def scale2d(s):
+    """Homogeneous 2D scaling."""
+    if isinstance(s, (int, float)):
+        s = [s,s]        
+    S = np.array([
+        [s[0], 0,    0],
+        [0,    s[1], 0],
+        [0,    0,    1]
+    ])
+    return S
+
+
+def rotation2d(alpha):
+    ca = np.cos(alpha)
+    sa = np.sin(alpha)
+    R = np.array([
+        [ca, -sa, 0],
+        [sa,  ca, 0],
+        [ 0,   0, 1]
+    ])
+    return R
+
+
+def rigid_transform(alpha, tu, tv):
+    """Homogeneous 2D rigid transformation."""
+    ca = np.cos(alpha)
+    sa = np.sin(alpha)
+    R = np.array([
+        [ca, -sa, tu],
+        [sa,  ca, tv],
+        [ 0,   0,  1]
+    ])
+    return R
+    
+    
+def translation(t):
+    """Homogeneous 3D translation."""
+    T = np.array([
+        [1, 0, 0, t[0]],
+        [0, 1, 0, t[1]],
+        [0, 0, 1, t[2]],
+        [0, 0, 0, 1]
+    ])
+    return T
+    
+
+def scale(s):
+    """Homogeneous 3D scaling."""
+    if isinstance(s, (int, float)):
+        s = [s,s,s]
+    S = np.array([
+        [s[0], 0,    0,    0],
+        [0,    s[1], 0,    0],
+        [0,    0,    s[2], 0],
+        [0,    0,    0,    1]
+    ])
+    return S
+    
+    
+def rotation_x(alpha):
+    """Homogeneous rotation about X-axis."""
+    ca = np.cos(alpha)
+    sa = np.sin(alpha)
+    R = np.array([
+        [1,  0,   0, 0],
+        [0, ca, -sa, 0],
+        [0, sa,  ca, 0],
+        [0,  0,   0, 1]
+    ])
+    return R
+
+
+def rotation_y(alpha):
+    """Homogeneous rotation about Y-axis."""
+    ca = np.cos(alpha)
+    sa = np.sin(alpha)
+    R = np.array([
+        [ca,  0, sa, 0],
+        [ 0,  1,  0, 0],
+        [-sa, 0, ca, 0],
+        [ 0,  0,  0, 1]
+    ])
+    return R
+
+
+def rotation_z(alpha):
+    """Homogeneous rotation about Z-axis."""
+    ca = np.cos(alpha)
+    sa = np.sin(alpha)
+    R = np.array([
+        [ca, -sa, 0, 0],
+        [sa,  ca, 0, 0],
+        [ 0,   0, 1, 0],
+        [ 0,   0, 0, 1]
+    ])
+    return R
+
+
+def euler_to_rotation_matrix(euler_angles):
+    """Rotation from Euler angles. Assumes input is a 3-vector containing the Euler angles."""
+    rotation_x_matrix = rotation_x(euler_angles[0])
+    rotation_y_matrix = rotation_y(euler_angles[1])
+    rotation_z_matrix = rotation_z(euler_angles[2])
+
+    R = rotation_x_matrix @ rotation_y_matrix @ rotation_z_matrix
+    return R
+
+
+def euler_from_rotation_matrix(R):
+    """Euler angles from rotation matrix.
+    Assumes the top-left 3x3 submatrix of R is in fact an orthogonal matrix."""
+    ry = np.arcsin(R[0, 2])
+    rz = np.arccos(R[0, 0] / np.cos(ry))
+    rx = np.arccos(R[2, 2] / np.cos(ry))
+    euler_angles = np.array([rx, ry, rz])
+    return euler_angles
+

ProjectiveGeometry23/pluecker.py

@@ -0,0 +1,161 @@
+"""
+Representation of lines in 3D space in Plücker coordinates.
+Author: André Aichert
+Date: June 22, 2023
+"""
+
+import numpy as np
+from .utils import dot, cvec, append
+from .utils import homogenize, dehomogenize, RP3Point
+from scipy.optimize import minimize
+
+
+def join_points(A, B):
+    """Join two points to form a 3D line."""
+    A = A.flatten()
+    B = B.flatten()
+    return cvec([
+            A[0] * B[1] - A[1] * B[0],
+            A[0] * B[2] - A[2] * B[0],
+            A[0] * B[3] - A[3] * B[0],
+            A[1] * B[2] - A[2] * B[1],
+            A[1] * B[3] - A[3] * B[1],
+            A[2] * B[3] - A[3] * B[2]
+        ])
+
+
+def meet_planes(P, Q):
+    """Intersect two planes to form a 3D line."""
+    P = P.flatten()
+    Q = Q.flatten()
+    return cvec([
+            P[2] * Q[3] - P[3] * Q[2],
+            P[3] * Q[1] - P[1] * Q[3],
+            P[1] * Q[2] - P[2] * Q[1],
+            P[0] * Q[3] - P[3] * Q[0],
+            P[2] * Q[0] - P[0] * Q[2],
+            P[0] * Q[1] - P[1] * Q[0]
+        ])
+    
+
+def join(L, X):
+    """Join a line L and a point X to form a plane."""
+    L = L.flatten()
+    X = X.flatten()
+    return cvec([
+                          + X[1] * L[5] - X[2] * L[4] + X[3] * L[3],  # noqa
+            - X[0] * L[5]               + X[2] * L[2] - X[3] * L[1],  # noqa
+            + X[0] * L[4] - X[1] * L[2]               + X[3] * L[0],  # noqa
+            - X[0] * L[3] + X[1] * L[1] - X[2] * L[0]                 # noqa
+        ])
+    
+
+def meet(L, E):
+    """Intersect a line L and a plane E to form a point."""
+    L = L.flatten()
+    E = E.flatten()
+    return cvec([
+                          - E[1] * L[0] - E[2] * L[1] - E[3] * L[2],  # noqa
+            + E[0] * L[0]               - E[2] * L[3] - E[3] * L[4],  # noqa
+            + E[0] * L[1] + E[1] * L[3]               - E[3] * L[5],  # noqa
+            + E[0] * L[2] + E[1] * L[4] + E[2] * L[5]                 # noqa
+        ])
+    
+
+def direction(L):
+    """Direction of a 3D line in Plücker coordinates.
+    Note that for dual Plücker coordinates, the functions direction and moment will be swapped."""
+    return np.array([-L[2][0], -L[4][0], -L[5][0]]).reshape(-1, 1)
+    
+
+def moment(L):
+    """Moment of a 3D line in Plücker coordinates.
+    Note that for dual Plücker coordinates, the functions direction and moment will be swapped."""
+    return np.array([L[3][0], -L[1][0], L[0][0]]).reshape(-1, 1)
+    
+
+def closest_point_to_origin(L):
+    """Closest point on line L to the origin."""
+    L = L.flatten()
+    return np.array([
+         L[4] * L[0] + L[1] * L[5],  # noqa
+        -L[0] * L[2] + L[3] * L[5],  # noqa
+        -L[1] * L[2] - L[3] * L[4],  # noqa
+        -L[2] * L[2] - L[4] * L[4] - L[5] * L[5] # noqa
+    ]).reshape(-1, 1)
+
+
+def closest_to_point(L, X):
+    """Compute the closest point on the line L to a given point X."""
+    d = direction(L)
+    plane_through_X_orthogonal_to_L = RP3Point(d[0][0], d[1][0], d[2][0], -dot(d, dehomogenize(X)))
+    closest_point_to_X_on_L = meet(L, plane_through_X_orthogonal_to_L)
+    return closest_point_to_X_on_L
+
+
+def distance_to_point(L, X):
+    """Compute the distance of a line L to a given point X."""
+    I = closest_to_point(L, X)
+    return np.linalg.norm(dehomogenize(I) - dehomogenize(X))
+
+
+def point_closest_to_lines(lines, eps=1e-9):
+    """Find point with minimal distance to a line bundle and return (distance, point).
+    Returns intersection of two or more lines, if they do intersect."""
+    def mean_distance(Xe):
+        """Function to be minimized. Takes 3D *euclidian* coordinates."""
+        return np.sum([distance_to_point(L, RP3Point(Xe[0], Xe[1] ,Xe[2], 1)) for L in lines]) / len(lines)
+    result = minimize(mean_distance, [0, 0, 0], tol=eps)
+    return result.fun, homogenize(result.x)
+
+
+def distance_to_origin(L):
+    """Distance of a line to the origin."""
+    m = moment(L)
+    d = direction(L)
+    return np.linalg.norm(m) / np.linalg.norm(d)
+
+
+def matrixDual(L):
+    """Anti-symmetric matrix for the join operation using dual Plücker coordinates."""
+    L = L.flatten()
+    B = np.array([
+        [    0,  L[5], -L[4],  L[3]],  # noqa
+        [-L[5],     0,  L[2], -L[1]],  # noqa
+        [ L[4], -L[2],     0,  L[0]],  # noqa
+        [-L[3],  L[1], -L[0],     0]   # noqa
+    ])
+    return B
+
+
+def matrix(L):
+    """Anti-symmetric matrix for the meet operation using Plücker coordinates."""
+    L = L.flatten()
+    B = np.array([
+        [   0, -L[0], -L[1], -L[2]],  # noqa
+        [L[0],     0, -L[3], -L[4]],  # noqa
+        [L[1],  L[3],     0, -L[5]],  # noqa
+        [L[2],  L[4],  L[5],     0]   # noqa
+    ])
+    return B
+
+
+def projection_matrix(P):
+    """Compute Sturm-style projection matrix for central projection of Plücker lines. Projection from Plücker coordinates directly to 2D lines written in a single 3x6 matrix. P is a standard 3x4 projection matrix."""
+    PL = np.array([
+        [P[1, 0]*P[2, 1]-P[1, 1]*P[2, 0],  P[1, 0]*P[2, 2]-P[1, 2]*P[2, 0],  P[1, 0]*P[2, 3]-P[1, 3]*P[2, 0],  P[1, 1]*P[2, 2]-P[1, 2]*P[2, 1],  P[1, 1]*P[2, 3]-P[1, 3]*P[2, 1],  P[1, 2]*P[2, 3]-P[1, 3]*P[2, 2]],  # noqa
+        [P[0, 1]*P[2, 0]-P[0, 0]*P[2, 1], -P[0, 0]*P[2, 2]+P[0, 2]*P[2, 0], -P[0, 0]*P[2, 3]+P[0, 3]*P[2, 0], -P[0, 1]*P[2, 2]+P[0, 2]*P[2, 1], -P[0, 1]*P[2, 3]+P[0, 3]*P[2, 1], -P[0, 2]*P[2, 3]+P[0, 3]*P[2, 2]],  # noqa
+        [P[0, 0]*P[1, 1]-P[0, 1]*P[1, 0],  P[0, 0]*P[1, 2]-P[0, 2]*P[1, 0],  P[0, 0]*P[1, 3]-P[0, 3]*P[1, 0],  P[0, 1]*P[1, 2]-P[0, 2]*P[1, 1],  P[0, 1]*P[1, 3]-P[0, 3]*P[1, 1],  P[0, 2]*P[1, 3]-P[0, 3]*P[1, 2]]   # noqa
+    ])
+    return PL
+
+
+def project(L, P):
+    """Directly project 3D line in Plücker coordinates to 2D line.
+    In Python, it may be faster to just do np.dot(projection_matrix(P), L)"""
+    line = np.array([
+        L[0]*(P[1, 0]*P[2, 1]-P[1, 1]*P[2, 0]) + L[1]*( P[1, 0]*P[2, 2]-P[1, 2]*P[2, 0]) + L[2]*( P[1, 0]*P[2, 3]-P[1, 3]*P[2, 0]) + L[3]*( P[1, 1]*P[2, 2]-P[1, 2]*P[2, 1]) + L[4]*( P[1, 1]*P[2, 3]-P[1, 3]*P[2, 1]) + L[5]*( P[1, 2]*P[2, 3]-P[1, 3]*P[2, 2]),  # noqa
+        L[0]*(P[0, 1]*P[2, 0]-P[0, 0]*P[2, 1]) + L[1]*(-P[0, 0]*P[2, 2]+P[0, 2]*P[2, 0]) + L[2]*(-P[0, 0]*P[2, 3]+P[0, 3]*P[2, 0]) + L[3]*(-P[0, 1]*P[2, 2]+P[0, 2]*P[2, 1]) + L[4]*(-P[0, 1]*P[2, 3]+P[0, 3]*P[2, 1]) + L[5]*(-P[0, 2]*P[2, 3]+P[0, 3]*P[2, 2]),  # noqa
+        L[0]*(P[0, 0]*P[1, 1]-P[0, 1]*P[1, 0]) + L[1]*( P[0, 0]*P[1, 2]-P[0, 2]*P[1, 0]) + L[2]*( P[0, 0]*P[1, 3]-P[0, 3]*P[1, 0]) + L[3]*( P[0, 1]*P[1, 2]-P[0, 2]*P[1, 1]) + L[4]*( P[0, 1]*P[1, 3]-P[0, 3]*P[1, 1]) + L[5]*( P[0, 2]*P[1, 3]-P[0, 3]*P[1, 2])   # noqa
+    ])
+    return line