dtype=np.float32)
H10 = np.linalg.inv(H01)
# The feature I'm going to test with. This is the corner of one of the towers
q0 = np.array((294, 159), dtype=np.float32)
# The transformed image. The matcher should end-up reversing this
# transformation, since it will be given the homography.
#
# shape (H,W,2)
image1 = \
mrcal.transform_image(
image,
mrcal.apply_homography(
H01,
nps.glue(*[ nps.dummy(arr, -1) for arr in \
np.meshgrid( np.arange(500),
np.arange(600))],
axis=-1).astype(np.float32) ))
# I have the source images and the "right" homography and the "right" matching
# pixels coords. Run the matcher, and compare
templatesize = (30, 20)
search_radius = 50
H10_shifted = H10.copy()
H10_shifted[0, 2] += 10.2
H10_shifted[1, 2] -= 20.4
q1_matched, diagnostics = \
pdf='pdf size 8in,6in noenhanced solid color font ",12"',
svg='svg size 800,600 noenhanced solid dynamic font ",14"',
png='pngcairo size 1024,768 transparent noenhanced crop font ",12"',
gp='gp')
for extension in terminal.keys():
mrcal.show_geometry(
list(models) + list(models_rectified),
("camera0", "camera1", "camera0_rectified", "camera1_rectified"),
show_calobjects=False,
_set=('xyplane at -0.5', 'view 60,30,1.7'),
terminal=terminal[extension],
hardcopy=f'/tmp/stereo-geometry-{kind}.{extension}')
# Generate the rectified images, and write to disk
images_rectified = [
mrcal.transform_image(images[i], rectification_maps[i])
for i in range(2)
]
cv2.imwrite(f'/tmp/rectified0-{kind}.jpg', images_rectified[0])
cv2.imwrite(f'/tmp/rectified1-{kind}.jpg', images_rectified[1])
# Perform stereo-matching with OpenCV to produce a disparity map, which we write
# to disk
block_size = 5
max_disp = 400
max_disp = 16 * round(max_disp / 16) # round to nearest multiple of 16
stereo = \
cv2.StereoSGBM_create(
minDisparity = 0,
numDisparities = max_disp,
blockSize = block_size,
def image_transformation_map(model_from, model_to,
use_rotation = False,
plane_n = None,
plane_d = None,
mask_valid_intrinsics_region_from = False):
r'''Compute a reprojection map between two models
SYNOPSIS
model_orig = mrcal.cameramodel("xxx.cameramodel")
image_orig = cv2.imread("image.jpg")
model_pinhole = mrcal.pinhole_model_for_reprojection(model_orig,
fit = "corners")
mapxy = mrcal.image_transformation_map(model_orig, model_pinhole)
image_undistorted = mrcal.transform_image(image_orig, mapxy)
# image_undistorted is now a pinhole-reprojected version of image_orig
Returns the transformation that describes a mapping
- from pixel coordinates of an image of a scene observed by model_to
- to pixel coordinates of an image of the same scene observed by model_from
This transformation can then be applied to a whole image by calling
mrcal.transform_image().
This function returns a transformation map in an (Nheight,Nwidth,2) array. The
image made by model_to will have shape (Nheight,Nwidth). Each pixel (x,y) in
this image corresponds to a pixel mapxy[y,x,:] in the image made by model_from.
This function has 3 modes of operation:
- intrinsics-only
This is the default. Selected if
- use_rotation = False
- plane_n = None
- plane_d = None
All of the extrinsics are ignored. If the two cameras have the same
orientation, then their observations of infinitely-far-away objects will line
up exactly
- rotation
This can be selected explicitly with
- use_rotation = True
- plane_n = None
- plane_d = None
Here we use the rotation component of the relative extrinsics. The relative
translation is impossible to use without knowing what we're looking at, so IT
IS ALWAYS IGNORED. If the relative orientation in the models matches reality,
then the two cameras' observations of infinitely-far-away objects will line up
exactly
- plane
This is selected if
- use_rotation = True
- plane_n is not None
- plane_d is not None
We map observations of a given plane in camera FROM coordinates
coordinates to where this plane would be observed by camera TO. This uses
ALL the intrinsics, extrinsics and the plane representation. If all of
these are correct, the observations of this plane would line up exactly in
the remapped-camera-fromimage and the camera-to image. The plane is
represented in camera-from coordinates by a normal vector plane_n, and the
distance to the normal plane_d. The plane is all points p such that
inner(p,plane_n) = plane_d. plane_n does not need to be normalized; any
scaling is compensated in plane_d.
ARGUMENTS
- model_from: the mrcal.cameramodel object describing the camera used to capture
the input image
- model_to: the mrcal.cameramodel object describing the camera that would have
captured the image we're producing
- use_rotation: optional boolean, defaulting to False. If True: we respect the
relative rotation in the extrinsics of the camera models.
- plane_n: optional numpy array of shape (3,); None by default. If given, we
produce a transformation to map observations of a given plane to the same
pixels in the source and target images. This argument describes the normal
vector in the coordinate system of model_from. The plane is all points p such
that inner(p,plane_n) = plane_d. plane_n does not need to be normalized; any
scaling is compensated in plane_d. If given, plane_d should be given also, and
use_rotation should be True. if given, we use the full intrinsics and
extrinsics of both camera models
- plane_d: optional floating-point value; None by default. If given, we produce
a transformation to map observations of a given plane to the same pixels in
the source and target images. The plane is all points p such that
inner(p,plane_n) = plane_d. plane_n does not need to be normalized; any
scaling is compensated in plane_d. If given, plane_n should be given also, and
use_rotation should be True. if given, we use the full intrinsics and
extrinsics of both camera models
- mask_valid_intrinsics_region_from: optional boolean defaulting to False. If
True, points outside the valid-intrinsics region in the FROM image are set to
black, and thus do not appear in the output image
RETURNED VALUE
A numpy array of shape (Nheight,Nwidth,2) where Nheight and Nwidth represent the
imager dimensions of model_to. This array contains 32-bit floats, as required by
cv2.remap() (the function providing the internals of mrcal.transform_image()).
This array can be passed to mrcal.transform_image()
'''
if (plane_n is None and plane_d is not None) or \
(plane_n is not None and plane_d is None):
raise Exception("plane_n and plane_d should both be None or neither should be None")
if plane_n is not None and plane_d is not None and \
not use_rotation:
raise Exception("We're looking at remapping a plane (plane_d, plane_n are not None), so use_rotation should be True")
Rt_to_from = None
if use_rotation:
Rt_to_r = model_to. extrinsics_Rt_fromref()
Rt_r_from = model_from.extrinsics_Rt_toref()
Rt_to_from = mrcal.compose_Rt(Rt_to_r, Rt_r_from)
lensmodel_from,intrinsics_data_from = model_from.intrinsics()
lensmodel_to, intrinsics_data_to = model_to. intrinsics()
if re.match("LENSMODEL_OPENCV",lensmodel_from) and \
lensmodel_to == "LENSMODEL_PINHOLE" and \
plane_n is None and \
not mask_valid_intrinsics_region_from:
# This is a common special case. This branch works identically to the
# other path (the other side of this "if" can always be used instead),
# but the opencv-specific code is optimized and at one point ran faster
# than the code on the other side.
#
# The mask_valid_intrinsics_region_from isn't implemented in this path.
# It COULD be, then this faster path could be used
fxy_from = intrinsics_data_from[0:2]
cxy_from = intrinsics_data_from[2:4]
cameraMatrix_from = np.array(((fxy_from[0], 0, cxy_from[0]),
( 0, fxy_from[1], cxy_from[1]),
( 0, 0, 1)))
fxy_to = intrinsics_data_to[0:2]
cxy_to = intrinsics_data_to[2:4]
cameraMatrix_to = np.array(((fxy_to[0], 0, cxy_to[0]),
( 0, fxy_to[1], cxy_to[1]),
( 0, 0, 1)))
output_shape = model_to.imagersize()
distortion_coeffs = intrinsics_data_from[4: ]
if Rt_to_from is not None:
R_to_from = Rt_to_from[:3,:]
if np.trace(R_to_from) > 3. - 1e-12:
R_to_from = None # identity, so I pass None
else:
R_to_from = None
return nps.glue( *[ nps.dummy(arr,-1) for arr in \
cv2.initUndistortRectifyMap(cameraMatrix_from, distortion_coeffs,
R_to_from,
cameraMatrix_to, tuple(output_shape),
cv2.CV_32FC1)],
axis = -1)
W_from,H_from = model_from.imagersize()
W_to, H_to = model_to. imagersize()
# shape: (Nheight,Nwidth,2). Contains (x,y) rows
grid = np.ascontiguousarray(nps.mv(nps.cat(*np.meshgrid(np.arange(W_to),
np.arange(H_to))),
0,-1),
dtype = float)
if lensmodel_to == "LENSMODEL_PINHOLE":
# Faster path for the unproject. Nice, simple closed-form solution
fxy_to = intrinsics_data_to[0:2]
cxy_to = intrinsics_data_to[2:4]
v = np.zeros( (grid.shape[0], grid.shape[1], 3), dtype=float)
v[..., :2] = (grid-cxy_to)/fxy_to
v[..., 2] = 1
elif lensmodel_to == "LENSMODEL_STEREOGRAPHIC":
# Faster path for the unproject. Nice, simple closed-form solution
v = mrcal.unproject_stereographic(grid, *intrinsics_data_to[:4])
else:
v = mrcal.unproject(grid, lensmodel_to, intrinsics_data_to)
if plane_n is not None:
R_to_from = Rt_to_from[:3,:]
t_to_from = Rt_to_from[ 3,:]
# The homography definition. Derived in many places. For instance in
# "Motion and structure from motion in a piecewise planar environment"
# by Olivier Faugeras, F. Lustman.
A_to_from = plane_d * R_to_from + nps.outer(t_to_from, plane_n)
A_from_to = np.linalg.inv(A_to_from)
v = nps.matmult( v, nps.transpose(A_from_to) )
else:
if Rt_to_from is not None:
R_to_from = Rt_to_from[:3,:]
if np.trace(R_to_from) < 3. - 1e-12:
# rotation isn't identity. apply
v = nps.matmult(v, R_to_from)
mapxy = mrcal.project( v, lensmodel_from, intrinsics_data_from )
if mask_valid_intrinsics_region_from:
# Using matplotlib to compute the out-of-bounds points. It doesn't
# support broadcasting, so I do that manually with a clump/reshape
from matplotlib.path import Path
region = Path(model_from.valid_intrinsics_region())
is_inside = region.contains_points(nps.clump(mapxy,n=2)).reshape(mapxy.shape[:2])
mapxy[ ~is_inside, :] = -1
return mapxy.astype(np.float32)
def image_transformation_map(model_from,
model_to,
intrinsics_only=False,
distance=None,
plane_n=None,
plane_d=None,
mask_valid_intrinsics_region_from=False):
r'''Compute a reprojection map between two models
SYNOPSIS
model_orig = mrcal.cameramodel("xxx.cameramodel")
image_orig = cv2.imread("image.jpg")
model_pinhole = mrcal.pinhole_model_for_reprojection(model_orig,
fit = "corners")
mapxy = mrcal.image_transformation_map(model_orig, model_pinhole,
intrinsics_only = True)
image_undistorted = mrcal.transform_image(image_orig, mapxy)
# image_undistorted is now a pinhole-reprojected version of image_orig
Returns the transformation that describes a mapping
- from pixel coordinates of an image of a scene observed by model_to
- to pixel coordinates of an image of the same scene observed by model_from
This transformation can then be applied to a whole image by calling
mrcal.transform_image().
This function returns a transformation map in an (Nheight,Nwidth,2) array. The
image made by model_to will have shape (Nheight,Nwidth). Each pixel (x,y) in
this image corresponds to a pixel mapxy[y,x,:] in the image made by model_from.
One application of this function is to validate the models in a stereo pair. For
instance, reprojecting one camera's image at distance=infinity should produce
exactly the same image that is observed by the other camera when looking at very
far objects, IF the intrinsics and rotation are correct. If the images don't
line up well, we know that some part of the models is off. Similarly, we can use
big planes (such as observations of the ground) and plane_n, plane_d to validate.
This function has several modes of operation:
- intrinsics, extrinsics
Used if not intrinsics_only and \
plane_n is None and \
plane_d is None
This is the default. For each pixel in the output, we use the full model to
unproject a given distance out, and then use the full model to project back
into the other camera.
- intrinsics only
Used if intrinsics_only and \
plane_n is None and \
plane_d is None
Similar, but the extrinsics are ignored. We unproject the pixels in one model,
and project the into the other camera. The two camera coordinate systems are
assumed to line up perfectly
- plane
Used if plane_n is not None and
plane_d is not None
We map observations of a given plane in camera FROM coordinates coordinates to
where this plane would be observed by camera TO. We unproject each pixel in
one camera, compute the intersection point with the plane, and project that
intersection point back to the other camera. This uses ALL the intrinsics,
extrinsics and the plane representation. The plane is represented by a normal
vector plane_n, and the distance to the normal plane_d. The plane is all
points p such that inner(p,plane_n) = plane_d. plane_n does not need to be
normalized; any scaling is compensated in plane_d.
ARGUMENTS
- model_from: the mrcal.cameramodel object describing the camera used to capture
the input image. We always use the intrinsics. if not intrinsics_only: we use
the extrinsics also
- model_to: the mrcal.cameramodel object describing the camera that would have
captured the image we're producing. We always use the intrinsics. if not
intrinsics_only: we use the extrinsics also
- intrinsics_only: optional boolean, defaulting to False. If False: we respect
the relative transformation in the extrinsics of the camera models.
- distance: optional value, defaulting to None. Used only if not
intrinsics_only. We reproject points in space a given distance out. If
distance is None (the default), we look out to infinity. This is equivalent to
using only the rotation component of the extrinsics, ignoring the translation.
- plane_n: optional numpy array of shape (3,); None by default. If given, we
produce a transformation to map observations of a given plane to the same
pixels in the source and target images. This argument describes the normal
vector in the coordinate system of model_from. The plane is all points p such
that inner(p,plane_n) = plane_d. plane_n does not need to be normalized; any
scaling is compensated in plane_d. If given, plane_d should be given also, and
intrinsics_only should be False. if given, we use the full intrinsics and
extrinsics of both camera models
- plane_d: optional floating-point value; None by default. If given, we produce
a transformation to map observations of a given plane to the same pixels in
the source and target images. The plane is all points p such that
inner(p,plane_n) = plane_d. plane_n does not need to be normalized; any
scaling is compensated in plane_d. If given, plane_n should be given also, and
intrinsics_only should be False. if given, we use the full intrinsics and
extrinsics of both camera models
- mask_valid_intrinsics_region_from: optional boolean defaulting to False. If
True, points outside the valid-intrinsics region in the FROM image are set to
black, and thus do not appear in the output image
RETURNED VALUE
A numpy array of shape (Nheight,Nwidth,2) where Nheight and Nwidth represent the
imager dimensions of model_to. This array contains 32-bit floats, as required by
cv2.remap() (the function providing the internals of mrcal.transform_image()).
This array can be passed to mrcal.transform_image()
'''
if (plane_n is None and plane_d is not None) or \
(plane_n is not None and plane_d is None):
raise Exception(
"plane_n and plane_d should both be None or neither should be None"
)
if plane_n is not None and \
intrinsics_only:
raise Exception(
"We're looking at remapping a plane (plane_d, plane_n are not None), so intrinsics_only should be False"
)
if distance is not None and \
(plane_n is not None or intrinsics_only):
raise Exception(
"'distance' makes sense only without plane_n/plane_d and without intrinsics_only"
)
if intrinsics_only:
Rt_to_from = None
else:
Rt_to_from = mrcal.compose_Rt(model_to.extrinsics_Rt_fromref(),
model_from.extrinsics_Rt_toref())
lensmodel_from, intrinsics_data_from = model_from.intrinsics()
lensmodel_to, intrinsics_data_to = model_to.intrinsics()
if re.match("LENSMODEL_OPENCV",lensmodel_from) and \
lensmodel_to == "LENSMODEL_PINHOLE" and \
plane_n is None and \
not mask_valid_intrinsics_region_from and \
distance is None:
# This is a common special case. This branch works identically to the
# other path (the other side of this "if" can always be used instead),
# but the opencv-specific code is optimized and at one point ran faster
# than the code on the other side.
#
# The mask_valid_intrinsics_region_from isn't implemented in this path.
# It COULD be, then this faster path could be used
import cv2
fxy_from = intrinsics_data_from[0:2]
cxy_from = intrinsics_data_from[2:4]
cameraMatrix_from = np.array(
((fxy_from[0], 0, cxy_from[0]), (0, fxy_from[1], cxy_from[1]),
(0, 0, 1)))
fxy_to = intrinsics_data_to[0:2]
cxy_to = intrinsics_data_to[2:4]
cameraMatrix_to = np.array(
((fxy_to[0], 0, cxy_to[0]), (0, fxy_to[1], cxy_to[1]), (0, 0, 1)))
output_shape = model_to.imagersize()
distortion_coeffs = intrinsics_data_from[4:]
if Rt_to_from is not None:
R_to_from = Rt_to_from[:3, :]
if np.trace(R_to_from) > 3. - 1e-12:
R_to_from = None # identity, so I pass None
else:
R_to_from = None
return nps.glue( *[ nps.dummy(arr,-1) for arr in \
cv2.initUndistortRectifyMap(cameraMatrix_from, distortion_coeffs,
R_to_from,
cameraMatrix_to, tuple(output_shape),
cv2.CV_32FC1)],
axis = -1)
W_from, H_from = model_from.imagersize()
W_to, H_to = model_to.imagersize()
# shape: (Nheight,Nwidth,2). Contains (x,y) rows
grid = np.ascontiguousarray(nps.mv(
nps.cat(*np.meshgrid(np.arange(W_to), np.arange(H_to))), 0, -1),
dtype=float)
v = mrcal.unproject(grid, lensmodel_to, intrinsics_data_to)
if plane_n is not None:
R_to_from = Rt_to_from[:3, :]
t_to_from = Rt_to_from[3, :]
# The homography definition. Derived in many places. For instance in
# "Motion and structure from motion in a piecewise planar environment"
# by Olivier Faugeras, F. Lustman.
A_to_from = plane_d * R_to_from + nps.outer(t_to_from, plane_n)
A_from_to = np.linalg.inv(A_to_from)
v = nps.matmult(v, nps.transpose(A_from_to))
else:
if Rt_to_from is not None:
if distance is not None:
v = mrcal.transform_point_Rt(
mrcal.invert_Rt(Rt_to_from),
v / nps.dummy(nps.mag(v), -1) * distance)
else:
R_to_from = Rt_to_from[:3, :]
v = nps.matmult(v, R_to_from)
mapxy = mrcal.project(v, lensmodel_from, intrinsics_data_from)
if mask_valid_intrinsics_region_from:
# Using matplotlib to compute the out-of-bounds points. It doesn't
# support broadcasting, so I do that manually with a clump/reshape
from matplotlib.path import Path
region = Path(model_from.valid_intrinsics_region())
is_inside = region.contains_points(nps.clump(mapxy, n=2)).reshape(
mapxy.shape[:2])
mapxy[~is_inside, :] = -1
return mapxy.astype(np.float32)
print(f"Usage: {sys.argv[0]} model image yaw_deg what", file=sys.stderr)
sys.exit(1)
# I want a pinhole model to cover the middle 1/3rd of my pixels
W, H = model.imagersize()
fit_points = \
np.array((( W/3., H/3.),
( W*2./3., H/3.),
( W/3., H*2./3.),
( W*2./3., H*2./3.)))
model_pinhole = \
mrcal.pinhole_model_for_reprojection(model,
fit = fit_points,
scale_image = 0.5)
# yaw transformation: pure rotation around the y axis
rt_yaw = np.array((0., yaw_deg * np.pi / 180., 0, 0, 0, 0))
# apply the extra yaw transformation to my extrinsics
model_pinhole.extrinsics_rt_toref( \
mrcal.compose_rt(model_pinhole.extrinsics_rt_toref(),
rt_yaw) )
mapxy = mrcal.image_transformation_map(model, model_pinhole, use_rotation=True)
image_transformed = mrcal.transform_image(image, mapxy)
cv2.imwrite(f'/tmp/narrow-{what}.jpg', image_transformed)
model_pinhole.write(f'/tmp/pinhole-narrow-yawed-{what}.cameramodel')