def extract_chessboards(img, include_unrefined=False, use_corner_thresholding=True):
if img.shape[1] > 4000:
scale_factor = .2
elif img.shape[1] > 2000:
scale_factor = .5
else:
scale_factor = 1
if scale_factor < 1:
img_scaled = imresize(img, scale_factor, interp='bicubic')
else:
img_scaled = img
corners, v1, v2 = find_corners(img_scaled, use_corner_thresholding=use_corner_thresholding)
chessboards = chessboards_from_corners(corners, v1, v2)
chessboards = fix_orientations(chessboards, corners, img_scaled)
if scale_factor < 1:
corners = np.round(corners * (1.0/scale_factor)).astype(np.int)
img = prepare_image(img)
du, dv, angle, weight = c_pycb.gradient(img)
refined, v1, v2 = do_refine_corners(du, dv, angle, weight, corners, 15)
else:
refined = corners
if include_unrefined:
return refined, chessboards, corners
else:
return refined, chessboards
def extract_chessboards(img,
include_unrefined=False,
use_corner_thresholding=True):
if img.shape[1] > 4000:
scale_factor = .2
elif img.shape[1] > 2000:
scale_factor = .5
else:
scale_factor = 1
if scale_factor < 1:
img_scaled = imresize(img, scale_factor, interp='bicubic')
else:
img_scaled = img
corners, v1, v2 = find_corners(
img_scaled, use_corner_thresholding=use_corner_thresholding)
chessboards = chessboards_from_corners(corners, v1, v2)
chessboards = fix_orientations(chessboards, corners, img_scaled)
if scale_factor < 1:
corners = np.round(corners * (1.0 / scale_factor)).astype(np.int)
img = prepare_image(img)
du, dv, angle, weight = c_pycb.gradient(img)
refined, v1, v2 = do_refine_corners(du, dv, angle, weight, corners, 15)
else:
refined = corners
if include_unrefined:
return refined, chessboards, corners
else:
return refined, chessboards
def find_corners(img, tau=0.001, refine_corners=True, use_corner_thresholding=False):
img = prepare_image(img)
# Use a shared variable so we aren't copying the image each time we use it
#img_theano = theano.shared(img)
height, width = img.shape
du, dv, angle, weight = c_pycb.gradient(img)#, img_theano)
# scale input image
min = np.min(img)
max = np.max(img)
img = (img-min)/(max-min)
radius = [4, 8, 12]
# template properties
template_props = [
[ 0, pi/2, radius[0]],
[pi/4, -pi/4, radius[0]],
[ 0, pi/2, radius[1]],
[pi/4, -pi/4, radius[1]],
[ 0, pi/2, radius[2]],
[pi/4, -pi/4, radius[2]],
]
# Filter image
corners_img = np.zeros(img.shape)
for template_params in template_props:
template = c_pycb.Template(*template_params)
## Batch filters for speed
#filters = np.empty((4, template.a1.shape[0], template.a1.shape[1]),dtype=template.a1.dtype)
#filters[0,:,:] = template.a1
#filters[1,:,:] = template.a2
#filters[2,:,:] = template.b1
#filters[3,:,:] = template.b2
#s = theano_conv2d(img_theano, filters, border_mode='full').eval()
## Take off padding
#r = template_params[2]
#corners_a1_2 = s[0, r:height+r,r:width+r]
#corners_a2_2 = s[1, r:height+r,r:width+r]
#corners_b1_2 = s[2, r:height+r,r:width+r]
#corners_b2_2 = s[3, r:height+r,r:width+r]
#template = c_pycb.Template(*template_params)
vectorize = True
if vectorize:
cur_func = None
if template.a1.shape[0] == 9:
cur_func = tconv_9
elif template.a1.shape[0] == 17:
cur_func = tconv_17
elif template.a1.shape[0] == 25:
cur_func = tconv_25
template.a1 = template.a1[None, :][None, :]
template.a2 = template.a2[None, :][None, :]
template.b1 = template.b1[None, :][None, :]
template.b2 = template.b2[None, :][None, :]
filts = np.concatenate((template.a1, template.a2,
template.b1, template.b2), axis=0).astype("float32")
valz = cur_func(img.astype("float32")[None, :][None, :], filts)
corners_a1 = valz[0, 0]
corners_a2 = valz[0, 1]
corners_b1 = valz[0, 2]
corners_b2 = valz[0, 3]
else:
corners_a1, corners_a2, corners_b1, corners_b2 = c_pycb.conv_template(
img, template.a1, template.a2, template.b1, template.b2)
# print " " + str(template.a1.shape)
# print " " + str(template.a2.shape)
# print " " + str(template.b1.shape)
# print " " + str(template.b2.shape)
#corners_a1 = c_pycb.conv2(img, template.a1)
#corners_a2 = c_pycb.conv2(img, template.a2)
#corners_b1 = c_pycb.conv2(img, template.b1)
#corners_b2 = c_pycb.conv2(img, template.b2)
#corners_a1_3 = convolve2d(img, template.a1, mode='same')
#import pdb; pdb.set_trace()
#print corners_a1, corners_a2, corners_b1, corners_b2
#corners_a1 = convolve2d(img, template.a1, mode='same')
#corners_a2 = convolve2d(img, template.a2, mode='same')
#corners_b1 = convolve2d(img, template.b1, mode='same')
#corners_b2 = convolve2d(img, template.b2, mode='same')
# Compute mean
corners_mu = (corners_a1 + corners_a2 + corners_b1 + corners_b2)/4
# case 1: a=white, b=black
corners_a = np.minimum(corners_a1 - corners_mu,corners_a2 - corners_mu)
corners_b = np.minimum(corners_mu - corners_b1,corners_mu - corners_b2)
corners_1 = np.minimum(corners_a,corners_b)
# case 2: b=white, a=black
corners_a = np.minimum(corners_mu - corners_a1, corners_mu - corners_a2)
corners_b = np.minimum(corners_b1 - corners_mu, corners_b2 - corners_mu)
corners_2 = np.minimum(corners_a, corners_b)
# update corner map
corners_img = np.maximum(corners_img, corners_1)
corners_img = np.maximum(corners_img, corners_2)
# extract corner candidates via non maximum suppression
corners = c_pycb.non_maximum_supression(corners_img,3,0.025,5)
# subpixel refinement
if refine_corners:
corners, v1, v2 = do_refine_corners(du,dv,angle,weight,corners,10)
if len(corners) == 0:
return corners, v1, v2
# remove corners without edges
idx = np.where((v1[:,0]==0) & (v1[:,1]==0))[0]
corners = np.delete(corners, idx, 0)
v1 = np.delete(v1, idx, 0)
v2 = np.delete(v2, idx, 0)
if len(corners) == 0:
return corners, v1, v2
if use_corner_thresholding:
#% score corners
scores = score_corners(img,angle,weight,corners,v1,v2,radius)
# remove low scoring corners
idx = np.where(scores<tau)
corners = np.delete(corners, idx, 0)
v1 = np.delete(v1, idx, 0)
v2 = np.delete(v2, idx, 0)
scores = np.delete(scores, idx, 0)
# make v1(:,1)+v1(:,2) positive (=> comparable to c++ code)
idx = np.where((v1[:,0]+v1[:,1])<0)[0]
if len(idx) > 0:
v1[idx,:] = -v1[idx,:]
# make all coordinate systems right-handed (reduces matching ambiguities from 8 to 4)
corners_n1 = np.column_stack([v1[:,1], -v1[:,0]])
flip = -np.sign(corners_n1[:,0]*v2[:,0]+corners_n1[:,1]*v2[:,1])
v2 = v2*flip[:,np.newaxis].dot(np.ones((1,2)))
return corners, v1, v2
def find_corners(img,
tau=0.001,
refine_corners=True,
use_corner_thresholding=False):
img = prepare_image(img)
# Use a shared variable so we aren't copying the image each time we use it
#img_theano = theano.shared(img)
height, width = img.shape
du, dv, angle, weight = c_pycb.gradient(img) #, img_theano)
# scale input image
min = np.min(img)
max = np.max(img)
img = (img - min) / (max - min)
radius = [4, 8, 12]
# template properties
template_props = [
[0, pi / 2, radius[0]],
[pi / 4, -pi / 4, radius[0]],
[0, pi / 2, radius[1]],
[pi / 4, -pi / 4, radius[1]],
[0, pi / 2, radius[2]],
[pi / 4, -pi / 4, radius[2]],
]
# Filter image
corners_img = np.zeros(img.shape)
for template_params in template_props:
template = c_pycb.Template(*template_params)
## Batch filters for speed
#filters = np.empty((4, template.a1.shape[0], template.a1.shape[1]),dtype=template.a1.dtype)
#filters[0,:,:] = template.a1
#filters[1,:,:] = template.a2
#filters[2,:,:] = template.b1
#filters[3,:,:] = template.b2
#s = theano_conv2d(img_theano, filters, border_mode='full').eval()
## Take off padding
#r = template_params[2]
#corners_a1_2 = s[0, r:height+r,r:width+r]
#corners_a2_2 = s[1, r:height+r,r:width+r]
#corners_b1_2 = s[2, r:height+r,r:width+r]
#corners_b2_2 = s[3, r:height+r,r:width+r]
#template = c_pycb.Template(*template_params)
vectorize = True
if vectorize:
cur_func = None
if template.a1.shape[0] == 9:
cur_func = tconv_9
elif template.a1.shape[0] == 17:
cur_func = tconv_17
elif template.a1.shape[0] == 25:
cur_func = tconv_25
template.a1 = template.a1[None, :][None, :]
template.a2 = template.a2[None, :][None, :]
template.b1 = template.b1[None, :][None, :]
template.b2 = template.b2[None, :][None, :]
filts = np.concatenate(
(template.a1, template.a2, template.b1, template.b2),
axis=0).astype("float32")
valz = cur_func(img.astype("float32")[None, :][None, :], filts)
corners_a1 = valz[0, 0]
corners_a2 = valz[0, 1]
corners_b1 = valz[0, 2]
corners_b2 = valz[0, 3]
else:
corners_a1, corners_a2, corners_b1, corners_b2 = c_pycb.conv_template(
img, template.a1, template.a2, template.b1, template.b2)
# print " " + str(template.a1.shape)
# print " " + str(template.a2.shape)
# print " " + str(template.b1.shape)
# print " " + str(template.b2.shape)
#corners_a1 = c_pycb.conv2(img, template.a1)
#corners_a2 = c_pycb.conv2(img, template.a2)
#corners_b1 = c_pycb.conv2(img, template.b1)
#corners_b2 = c_pycb.conv2(img, template.b2)
#corners_a1_3 = convolve2d(img, template.a1, mode='same')
#import pdb; pdb.set_trace()
#print corners_a1, corners_a2, corners_b1, corners_b2
#corners_a1 = convolve2d(img, template.a1, mode='same')
#corners_a2 = convolve2d(img, template.a2, mode='same')
#corners_b1 = convolve2d(img, template.b1, mode='same')
#corners_b2 = convolve2d(img, template.b2, mode='same')
# Compute mean
corners_mu = (corners_a1 + corners_a2 + corners_b1 + corners_b2) / 4
# case 1: a=white, b=black
corners_a = np.minimum(corners_a1 - corners_mu,
corners_a2 - corners_mu)
corners_b = np.minimum(corners_mu - corners_b1,
corners_mu - corners_b2)
corners_1 = np.minimum(corners_a, corners_b)
# case 2: b=white, a=black
corners_a = np.minimum(corners_mu - corners_a1,
corners_mu - corners_a2)
corners_b = np.minimum(corners_b1 - corners_mu,
corners_b2 - corners_mu)
corners_2 = np.minimum(corners_a, corners_b)
# update corner map
corners_img = np.maximum(corners_img, corners_1)
corners_img = np.maximum(corners_img, corners_2)
# extract corner candidates via non maximum suppression
corners = c_pycb.non_maximum_supression(corners_img, 3, 0.025, 5)
# subpixel refinement
if refine_corners:
corners, v1, v2 = do_refine_corners(du, dv, angle, weight, corners, 10)
if len(corners) == 0:
return corners, v1, v2
# remove corners without edges
idx = np.where((v1[:, 0] == 0) & (v1[:, 1] == 0))[0]
corners = np.delete(corners, idx, 0)
v1 = np.delete(v1, idx, 0)
v2 = np.delete(v2, idx, 0)
if len(corners) == 0:
return corners, v1, v2
if use_corner_thresholding:
#% score corners
scores = score_corners(img, angle, weight, corners, v1, v2, radius)
# remove low scoring corners
idx = np.where(scores < tau)
corners = np.delete(corners, idx, 0)
v1 = np.delete(v1, idx, 0)
v2 = np.delete(v2, idx, 0)
scores = np.delete(scores, idx, 0)
# make v1(:,1)+v1(:,2) positive (=> comparable to c++ code)
idx = np.where((v1[:, 0] + v1[:, 1]) < 0)[0]
if len(idx) > 0:
v1[idx, :] = -v1[idx, :]
# make all coordinate systems right-handed (reduces matching ambiguities from 8 to 4)
corners_n1 = np.column_stack([v1[:, 1], -v1[:, 0]])
flip = -np.sign(corners_n1[:, 0] * v2[:, 0] + corners_n1[:, 1] * v2[:, 1])
v2 = v2 * flip[:, np.newaxis].dot(np.ones((1, 2)))
return corners, v1, v2