def affine_estimate(self, w, depth_reg=0.085, weights=None, scale=10.0,
scale_mean=0.0016 * 1.8 * 1.2, scale_std=1.2 * 0,
cap_scale=-0.00129):
"""
Quick switch to allow reconstruction at unknown scale returns a,r
and scale
"""
weights = np.zeros((0, 0, 0)) if weights is None else weights
s = np.empty((self.sigma.shape[0], self.sigma.shape[1] + 4)) # e,y,x,z
s[:, :4] = 10 ** -5 # Tiny but makes stuff well-posed
s[:, 0] = scale_std
s[:, 4:] = self.sigma
s[:, 4:-1] *= scale
e2 = np.zeros((self.e.shape[0], self.e.shape[
1] + 4, 3, self.e.shape[3]))
e2[:, 1, 0] = 1.0
e2[:, 2, 1] = 1.0
e2[:, 3, 0] = 1.0
# This makes the least_squares problem ill posed, as X,Z are
# interchangable
# Hence regularisation above to speed convergence and stop blow-up
e2[:, 0] = self.mu
e2[:, 4:] = self.e
t_m = np.zeros_like(self.mu)
res, a, r = pick_e(w, e2, t_m, self.cam, s, weights=weights,
interval=0.01, depth_reg=depth_reg,
scale_prior=scale_mean)
scale = a[:, :, 0]
reestimate = scale > cap_scale
m = self.mu * cap_scale
for i in range(scale.shape[0]):
if reestimate[i].sum() > 0:
ehat = e2[i:i + 1, 1:]
mhat = m[i:i + 1]
shat = s[i:i + 1, 1:]
(res2, a2, r2) = pick_e(
w[reestimate[i]], ehat, mhat, self.cam, shat,
weights=weights[reestimate[i]],
interval=0.01, depth_reg=depth_reg,
scale_prior=scale_mean
)
res[i:i + 1, reestimate[i]] = res2
a[i:i + 1, reestimate[i], 1:] = a2
a[i:i + 1, reestimate[i], 0] = cap_scale
r[i:i + 1, :, reestimate[i]] = r2
scale = a[:, :, 0]
a = a[:, :, 1:] / a[:, :, 0][:, :, np.newaxis]
return res, e2[:, 1:], a, r, scale
def affine_estimate(self, w, depth_reg=0.085, weights=None, scale=10.0,
scale_mean=0.0016 * 1.8 * 1.2, scale_std=1.2 * 0,
cap_scale=-0.00129):
"""
Quick switch to allow reconstruction at unknown scale returns a,r
and scale
"""
weights = np.zeros((0, 0, 0)) if weights is None else weights
s = np.empty((self.sigma.shape[0], self.sigma.shape[1] + 4)) # e,y,x,z
s[:, :4] = 10 ** -5 # Tiny but makes stuff well-posed
s[:, 0] = scale_std
s[:, 4:] = self.sigma
s[:, 4:-1] *= scale
e2 = np.zeros((self.e.shape[0], self.e.shape[
1] + 4, 3, self.e.shape[3]))
e2[:, 1, 0] = 1.0
e2[:, 2, 1] = 1.0
e2[:, 3, 0] = 1.0
# This makes the least_squares problem ill posed, as X,Z are
# interchangable
# Hence regularisation above to speed convergence and stop blow-up
e2[:, 0] = self.mu
e2[:, 4:] = self.e
t_m = np.zeros_like(self.mu)
res, a, r = pick_e(w, e2, t_m, self.cam, s, weights=weights,
interval=0.01, depth_reg=depth_reg,
scale_prior=scale_mean)
scale = a[:, :, 0]
reestimate = scale > cap_scale
m = self.mu * cap_scale
for i in range(scale.shape[0]):
if reestimate[i].sum() > 0:
ehat = e2[i:i + 1, 1:]
mhat = m[i:i + 1]
shat = s[i:i + 1, 1:]
(res2, a2, r2) = pick_e(
w[reestimate[i]], ehat, mhat, self.cam, shat,
weights=weights[reestimate[i]],
interval=0.01, depth_reg=depth_reg,
scale_prior=scale_mean
)
res[i:i + 1, reestimate[i]] = res2
a[i:i + 1, reestimate[i], 1:] = a2
a[i:i + 1, reestimate[i], 0] = cap_scale
r[i:i + 1, :, reestimate[i]] = r2
scale = a[:, :, 0]
a = a[:, :, 1:] / a[:, :, 0][:, :, np.newaxis]
return res, e2[:, 1:], a, r, scale