def best_into_mapping(domain, H, codomain):
"""
Associates each point in *domain* with the point in *codomain* nearest its image
under homography *H*. Returns an ordered list of images corresponding to points
in *domain*, along with some indication of how good or bad the match is.
"""
RANGE = []
total_error = 0
for preimage in domain:
modeled = H * preimage
min_error = None
image = None
for point in codomain:
error = sqdist(htoc(modeled), htoc(point))
if min_error == None or error < min_error:
min_error = error
image = point
total_error += min_error
RANGE.append(image)
return RANGE, total_error
def _reshape(points2, HOM_OUT=None):
"""
Returns a matrix containing two-dimensional representations of corner points in
*points2*, in a shape which hopefully corresponds to that of the calibration
board on which the corners are located. The associated homography is constrained
to map the corners of a pre-image board to the observed corners, and the
board dimensions must be co-divisors of len(*points2*). If requested,
np.savetxt()'s this homography to *HOM_OUT*.
"""
corners = cl.corners(points2)
npts = len(points2)
min_error = None
H = None
height = None
width = None
ordered = None
for (rows, cols) in [ (div, npts/div)
for div in range(2, npts/2 + 1)
if npts % div == 0 ]:
h = hm.homography(_box(rows, cols), [ ctoh(p) for p in corners ])
ordering, error = hm.best_into_mapping(_board(rows, cols),
h,
[ ctoh(p) for p in points2 ])
if min_error == None or error < min_error:
min_error = error
H = h
height = rows
width = cols
ordered = ordering
print " sum of squared error:", min_error
if HOM_OUT != None:
np.savetxt(HOM_OUT, H)
# return the homography and a matrix of the points in the shape of the board
return np.reshape(np.asarray([ htoc(p)
for p in ordered ]),
(height, width, -1))