def zl_predict(ls, dist_max, u0, v0, width, height, params):
hist = []
ang = []
for i in range(ls.shape[0]):
ang.append(line_angle2(ls[i, :]))
if ang[i] < -np.pi / 4:
ang[i] = ang[i] + np.pi
if np.abs(ang[i] - np.pi / 2) < params.theta_v:
l = line_hmg_from_two_points(ls[i, 0:2], ls[i, 2:4])
dist = np.abs(np.dot(l, np.array([width / 2, height / 2, 1])))
if dist < dist_max:
hist.append(ang[i])
[N, edges0] = np.histogram(
hist,
np.arange(np.pi / 2 - np.pi / 8, np.pi / 2 + np.pi / 8 + np.pi / 180,
np.pi / 180))
N[np.where(N <= 5)[0]] = 0
if np.sum(N) == 0:
N[int(np.round(len(N) / 2 + 0.1)) - 1] = 1
edges = (edges0[:-1] + edges0[1:]) / 2
max_modes = np.zeros([2 * len(N)])
H = np.zeros([len(N)])
Nout = mnf_modes.mnf(np.double(N), len(N), 1, max_modes, H)
max_modes = max_modes[:Nout * 2]
max_modes = max_modes.reshape(2, Nout).T
H = H[:Nout]
H = H.reshape(Nout)
if len(max_modes) == 0 or max_modes.shape[1] == 0:
max_modes = np.array([[1, N.shape[0]]])
H = 0
else:
I = np.argsort(-H)
H.sort()
H = H[::-1]
max_modes = max_modes[I, :]
zl = []
for i in range(max_modes.shape[0]):
Ni = np.zeros(N.shape[0])
a = max_modes[i, 0]
b = max_modes[i, 1]
Ni[int(a - 1):int(b)] = N[int(a - 1):int(b)]
m = np.max(Ni)
I = np.where(Ni == m)[0]
j = I[0]
a = edges[j]
l = np.abs(v0 / np.sin(a))
zl.append(u0 + l * np.cos(np.pi - a))
return zl
def vp_predict(lines_homo, initialIds, horizon_homo, params):
# compute intersections between the line segments and the HL candidate
inter_homo = np.cross(lines_homo[:, initialIds].T, horizon_homo).T
inter_homo = inter_homo / np.sqrt(np.sum(np.power(inter_homo, 2), axis=0))
inter_homo = inter_homo * np.sign(inter_homo[2, :] + np.finfo(float).eps)
inter_homo[2, :] = np.where(inter_homo[2, :] != 0, inter_homo[2, :],
np.finfo(float).eps)
inter_pts = inter_homo[:2, :] / inter_homo[2, :]
# compute the MMMs of the coordinate histogram
max_modes = []
a = horizon_homo[0]
b = horizon_homo[1]
c = horizon_homo[2]
A_hmg = np.cross(horizon_homo, np.array([b, -a, 0]))
A = np.array([A_hmg[0] / A_hmg[2], A_hmg[1] / A_hmg[2]])
rho = np.abs(c) / np.sqrt(np.power(a, 2) + np.power(b, 2))
rho2 = np.sqrt(inter_pts[0, :] * inter_pts[0, :] +
inter_pts[1, :] * inter_pts[1, :])
if rho > 1:
p = np.arccos(rho / rho2) / np.pi
else:
p = np.zeros(rho2.shape)
d = np.sqrt(np.abs(rho2 * rho2 - np.power(rho, 2)))
I = np.where(rho2 <= 1)[0]
if len(I) != 0:
p[I] = d[I] / np.pi
I = np.where(rho2 > 1)[0]
if len(I) != 0:
d2 = np.sqrt(rho2 * rho2 - 1)
beta = np.arctan(d2)
p[I] = (beta[I] + d[I] - d2[I]) / np.pi
tmp = inter_pts - np.array([A]).T.dot(np.ones([1, inter_pts.shape[1]]))
dt = np.array([b, -a]).dot(tmp)
I = np.where(dt < 0)[0]
p[I] = -p[I]
[N, edges] = np.histogram(p, params.L_vp)
# import scipy.io as sio
# tmp_mat = sio.loadmat('vp_Nedges.mat')
# [N, edges] = [tmp_mat['N'][0], tmp_mat['edges'][0]]
max_modes = np.zeros([2 * len(N)])
H = np.zeros([len(N)])
Nout = mnf_modes.mnf(np.double(N), len(N), 400, max_modes, H)
max_modes = max_modes[:Nout * 2]
max_modes = max_modes.reshape(2, Nout).T
H = H[:Nout]
H = H.reshape(Nout)
if len(max_modes) == 0:
max_modes = np.array([])
H = 0
else:
I = np.argsort(-H)
H.sort()
H = H[::-1]
max_modes = max_modes[I, :]
horgroups = []
scores = []
horvps_homo = []
for i in range(max_modes.shape[0]):
Ni = np.zeros([N.shape[0]])
a = max_modes[i, 0]
b = max_modes[i, 1]
Ni[int(a) - 1:int(b)] = N[int(a) - 1:int(b)]
m = np.max(Ni)
j = np.argmax(Ni)
p_i = (edges[j] + edges[j + 1]) / 2
vpId = np.argmin(np.abs(p - p_i))
horvps_homo.append(inter_homo[:, vpId])
scores.append(m)
edgesId = np.intersect1d(
np.where(p >= edges[int(a) - 1])[0],
np.where(p <= edges[int(b)])[0])
horgroups.append(edgesId)
if len(max_modes) == 0:
scores = [0]
sc = 0
horvps_homo = np.array([])
else:
#refine the VPs according to [Zhai et al.] and/or compute the scores
if params.hvp_refinement:
horvps_homo = vp_refinement(lines_homo, horvps_homo, horizon_homo,
params)
[scores, horgroups] = vp_score(horvps_homo, lines_homo,
params.score_function)
#sorted by score
sortIds = np.argsort(-scores)
scores.sort()
scores = scores[::-1]
horvps_homo = np.array(horvps_homo)[sortIds]
horgroups = horgroups[sortIds]
nvps = np.min([len(horgroups), 2])
sc = np.sum(scores[:nvps])
return [sc, horvps_homo, horgroups]
import mnf_modes import numpy as np histo = np.array([0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,8,15,20,19,8,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0], dtype=np.double) N = 45 epsilon = 1 outArray1 = np.zeros([2*N]) outArray2 = np.zeros([N]) print(len(outArray1)) Nout = mnf_modes.mnf(histo, N, epsilon,outArray1, outArray2) max_modes = outArray1[:Nout*2] max_modes = max_modes.reshape(2, Nout).T H = outArray2[:Nout] H = H.reshape(1, Nout).T print(max_modes) print(H)
def hl_predict(seglines, zenith_homo, u0, v0, width, height, focal, params):
L = params.L_h
x2 = zenith_homo[0] / zenith_homo[2]
y2 = zenith_homo[1] / zenith_homo[2]
tilt = np.arctan(x2 / y2)
if np.isnan(tilt):
tilt = 0
offsets = []
for i in range(seglines.shape[0]):
v = np.array([seglines[i, 2] - seglines[i, 0], seglines[i, 3] - seglines[i, 1]])
scal = np.dot(v/linalg.norm(v), np.array([np.cos(-tilt), np.sin(-tilt)]))
ang = np.arccos(scal)
if np.abs(ang) < params.theta_h * np.pi / 180:
tmp = np.dot(np.array([(seglines[i, 2] + seglines[i, 0])/2 - u0, (seglines[i, 3] + seglines[i, 1])/2 - v0]),
np.array([np.cos(-tilt + np.pi / 2), np.sin(-tilt + np.pi/2)]))
offsets.append(tmp)
offsets = np.asarray(offsets)
if len(offsets) == 0:
offsets = np.arange(-height/2, (height+1)/2)
[N, edges] = np.histogram(offsets, L)
# import scipy.io as sio
# tmp_mat = sio.loadmat('N_edges.mat')
# [N, edges] = [tmp_mat['N'][0], tmp_mat['edges'][0]]
max_modes = np.zeros([2 * len(N)])
H = np.zeros([len(N)])
Nout = mnf_modes.mnf(np.double(N), len(N), 1, max_modes, H)
max_modes = max_modes[:Nout * 2]
max_modes = max_modes.reshape(2, Nout).T
H = H[:Nout]
H = H.reshape(Nout)
# Modification 1
if Nout == 0:
max_modes = np.array([[1, N.shape[0]]])
H = np.array([-1])
else:
I = np.argsort(-H)
H.sort()
H = H[::-1]
max_modes = max_modes[I, :]
nmodes = max_modes.shape[0]
modes_offset = []
modes_left = []
modes_right = []
modes_homo = []
for i in range(nmodes):
Ni = np.zeros(N.shape[0])
a = max_modes[i, 0]
b = max_modes[i, 1]
Ni[int(a)-1 : int(b)] = N[int(a)-1 : int(b)]
bin = np.argmax(Ni)
#modes_offset.append(edges[0] + (bin + 1) / L * (edges[L-1] - edges[0]))
modes_offset.append(edges[0] + (bin + 1) / L * (edges[L] - edges[0]))
mnf_center = np.array([u0 + modes_offset[-1] * np.cos(-tilt + np.pi/2), v0 + modes_offset[-1] * np.sin(-tilt + np.pi/2)])
hmnf = line_hmg_from_two_points(mnf_center, mnf_center + np.array([np.cos(-tilt), np.sin(-tilt)]))
modes_left.append(-hmnf[2] / hmnf[1])
modes_right.append((-hmnf[0]*width -hmnf[2])/ hmnf[1])
mode_seg = np.array([[0, modes_left[0], width, modes_right[0]]])
modes_homo.append(np.squeeze(normalize.normalize(mode_seg, width, height, focal)))
modes_offset = np.array(modes_offset)
modes_left = np.array(modes_left)
modes_right = np.array(modes_right)
modes_homo = np.array(modes_homo).T
return [modes_homo, modes_offset, modes_left, modes_right, H]
def calculate_histogram(hvps_consensus_rectified, root, plot_redundant):
hvp_cum = np.vstack(hvps_consensus_rectified)
hvp_cum_pos = np.array([z_i if z_i[0] >= 0 else -z_i for z_i in hvp_cum])
hvp_degrees = np.degrees(np.arctan2(hvp_cum_pos[:, 0], hvp_cum_pos[:, 2]))
# hvp_periodic = np.where(hvp_degrees > 0, hvp_degrees, hvp_degrees+180)
hvp_periodic = hvp_degrees
hvp_periodic_2x = np.hstack([hvp_periodic, hvp_periodic + 180])
hvp_periodic_3x = np.hstack(
[hvp_periodic - 180, hvp_periodic, hvp_periodic + 180])
his_edge = 1
if plot_redundant:
plt.figure()
plt.hist(hvp_periodic, np.arange(181))
plt.title('histogram of angles (0-180)')
plt.xlabel('degree')
plt.ylabel('number')
plt.savefig(os.path.join(root, 'histogram.jpg'))
plt.figure()
plt.hist(hvp_periodic_2x, np.arange(361))
plt.title('histogram of angles (0-360)')
plt.xlabel('degree')
plt.ylabel('number')
plt.savefig(os.path.join(root, 'histogram_2x.jpg'))
plt.figure()
plt.hist(hvp_periodic_3x, np.arange(-181, 362))
plt.title('histogram of angles (-180-360)')
plt.xlabel('degree')
plt.ylabel('number')
plt.savefig(os.path.join(root, 'histogram_3x.jpg'))
if his_edge == 1:
hvp_histogram = np.histogram(hvp_periodic_3x, np.arange(-180, 361))
elif his_edge == 2:
hvp_histogram = np.histogram(hvp_periodic_3x, np.arange(-180, 361, 2))
elif his_edge == 3:
hvp_histogram = np.histogram(hvp_periodic_3x, np.arange(-180, 361, 3))
elif his_edge == 'auto':
hvp_histogram = np.histogram(hvp_periodic_3x, 'auto')
[N, edges] = hvp_histogram
max_modes = np.zeros([2 * len(N)])
H = np.zeros([len(N)])
his_epsilon = 1
Nout = mnf_modes.mnf(np.double(N), len(N), his_epsilon, max_modes, H)
max_modes = max_modes[:Nout * 2]
max_modes = max_modes.reshape(2, Nout).T
H = H[:Nout]
H = H.reshape(Nout)
if len(max_modes) == 0:
max_modes = np.array([])
H = 0
else:
I = np.argsort(-H)
H.sort()
H = H[::-1]
max_modes = max_modes[I, :]
horgroups = []
# scores = []
# horvps_homo = []
three_x_list = []
three_y_list = []
one_x_list = []
one_y_list = []
two_x_list = []
two_y_list = []
for i in range(max_modes.shape[0]):
Ni = np.zeros([N.shape[0]])
a = max_modes[i, 0]
b = max_modes[i, 1]
Ni[int(a) - 1:int(b)] = N[int(a) - 1:int(b)]
m = np.max(Ni)
j = np.argmax(Ni)
three_x_list.append(j - 180)
three_y_list.append(m)
# p_i = (edges[j] + edges[j+1]) / 2
# vpId = np.argmin(np.abs(p - p_i))
# horvps_homo.append(inter_homo[:,vpId])
# scores.append(m)
if 0 <= j - 180 / his_edge < 180 / his_edge:
#print('range{}-{}'.format(j - 180/his_edge, j - 180/his_edge + 1))
one_x_list.append(j - 180)
one_y_list.append(m)
two_x_list.append(j - 180)
two_y_list.append(m)
two_x_list.append(j - 180 + 180)
two_y_list.append(m)
############## Wrong shouldn't be a and b (fixed)
edgesId = np.intersect1d(np.where(hvp_periodic >= edges[j]),
np.where(hvp_periodic < edges[j + 1]))
horgroups.append(edgesId)
final_hvps_rectified = []
for i in range(len(horgroups)):
tmp = calculate_consensus_vps(hvp_cum_pos[horgroups[i]], 'svd')
# at most two vanishing points will be chosen
if i < 2:
final_hvps_rectified.append(tmp)
if plot_redundant:
# plot peaks on -180-360
plt.figure()
plt.hist(hvp_periodic_3x, np.arange(-181, 362))
plt.title('histogram of angles (-180-360)')
plt.scatter(three_x_list, three_y_list, s=20, c='r')
plt.xlabel('degree')
plt.ylabel('number')
plt.savefig(os.path.join(root, 'peaks_on_histogram_3x.jpg'))
# plot peaks on 0-180
plt.figure()
plt.hist(hvp_periodic, np.arange(181))
plt.title('histogram of angles (0-180)')
plt.scatter(one_x_list, one_y_list, s=20, c='r')
plt.xlabel('degree')
plt.ylabel('number')
plt.savefig(os.path.join(root, 'peaks_on_histogram.jpg'))
# plot peaks on 0-180
plt.figure()
plt.hist(hvp_periodic_2x, np.arange(361))
plt.title('histogram of angles (0-360)')
plt.scatter(two_x_list, two_y_list, s=20, c='r')
plt.xlabel('degree')
plt.ylabel('number')
plt.savefig(os.path.join(root, 'peaks_on_histogram_2x.jpg'))
return final_hvps_rectified