def edge_direction_3D(E, r=4):
D = conv_tri_3D(E, r)
Ox, Oy, Oz = np.gradient(D)
M = np.sqrt(Ox * Ox + Oy * Oy + Oz * Oz)
Ox, Oy, Oz = (Ox, Oy, Oz) / maxeps(M, 1e-5)
theta = np.arctan(Oy / maxeps(Oz, 1e-5))
phi = np.arctan(Oz / maxeps(M, 1e-5))
Px = np.sin(theta) * np.cos(phi)
Py = np.sin(theta) * np.sin(phi)
Pz = np.cos(phi)
return Px, Py, Pz, D
def edge_direction_3D(E,r=4):
D = conv_tri_3D(E,r)
Ox,Oy,Oz = np.gradient(D)
M = np.sqrt(Ox*Ox+Oy*Oy+Oz*Oz)
Ox,Oy,Oz = (Ox,Oy,Oz)/maxeps(M,1e-5)
theta = np.arctan(Oy/maxeps(Oz,1e-5))
phi = np.arctan(Oz/maxeps(M,1e-5))
Px= np.sin(theta)*np.cos(phi)
Py= np.sin(theta)*np.sin(phi)
Pz= np.cos(phi)
return Px,Py,Pz,D
def edge_direction_2D(E,r=5):
if(r<=1):
p=12/r/(r+2)-2;f=np.array([1,p,1])/(2+p); r=1;
else:
f = np.array(hstack((range(1,r),r+1,range(r,1,-1))),dtype=float)/(r+1)**2
F = np.zeros((len(f),len(f)))
F[:,len(f)/2] = f
F = signal.convolve2d(F,F.T,mode='full')
En = signal.convolve2d(E,F,mode='same')
Ox,Oy = np.gradient(En)
Oxx,_ = np.gradient(Ox);
Oxy,Oyy = np.gradient(Oy);
return np.arctan(Oyy*np.sign(-Oxy)/maxeps(Oxx,1e-5))
def edge_direction_2D(E, r=5):
if (r <= 1):
p = 12 / r / (r + 2) - 2
f = np.array([1, p, 1]) / (2 + p)
r = 1
else:
f = np.array(hstack((range(1, r), r + 1, range(r, 1, -1))),
dtype=float) / (r + 1)**2
F = np.zeros((len(f), len(f)))
F[:, len(f) / 2] = f
F = signal.convolve2d(F, F.T, mode='full')
En = signal.convolve2d(E, F, mode='same')
Ox, Oy = np.gradient(En)
Oxx, _ = np.gradient(Ox)
Oxy, Oyy = np.gradient(Oy)
return np.arctan(Oyy * np.sign(-Oxy) / maxeps(Oxx, 1e-5))
def eval_edgemap(E, GT, id=None, res_dir=None, K=99, thin=False):
thrs = np.linspace(1.0 / K, 1, K)
ret = list()
if id is None:
id = 'results'
if res_dir is not None:
fname = os.path.join(res_dir, str(id) + '.txt')
else:
fname = None
print("[", id, "]", end='', sep='')
if os.path.exists(fname):
ret_t = npa(np.loadtxt(fname))
print(fname, "exists! shape:", ret_t.shape)
return ret_t
print(K, E.shape, end='')
num_dots = 30
dot_on = np.ceil((1.0 + K) / num_dots)
for t, th in enumerate(thrs):
E1 = (normal_img(E) > maxeps(th)).astype(float)
_, _, matchE, matchG = correspondVoxels(
(E1 > 0).astype(float), (GT > 0).astype(float), .0075, 8)
matchE, matchG = matchE > 0, matchG > 0
d = dict(
th=th, #th
cntR=np.sum(matchG.astype(int)), #cntR
sumR=np.sum(GT.astype(int)), #sumR
cntP=np.sum(matchE.astype(int)), #cntP
sumP=np.sum(E1.astype(int)) #sumP
)
if t % dot_on == 0: print(".", end='')
ret.append(d)
print("!", end='')
if fname is not None:
keys = ['th', 'cntR', 'sumR', 'cntP', 'sumP']
ret2 = [[r[k] for k in keys] for r in ret]
header = "\t\t{:10s}{:10s}{:10s}{:10s}{:10s}".format(*keys)
np.savetxt(fname, ret2, fmt="%10g %10g %10g %10g %10g", header=header)
print("!")
return ret
def computeRPF(cntR, sumR, cntP, sumP):
R = np.divide(cntR.astype(float), maxeps(sumR))
P = np.divide(cntP.astype(float), maxeps(sumP))
d = npa(maxeps(P + R))
F = (2 * R * P) / d
return R, P, F