def silly_random_search(self, initial_trials):
## Estimate solution using RANSAC
bestv = np.Inf ## Smallest value found
args_f = (self.edgels, self.i_param, self.loss)
for k in range(initial_trials):
qtest = random_quaternion().canonical()
newv = corisco_aux.angle_error(qtest.q, *args_f)
## If the value is the best yet, store solution.
if newv <= bestv:
bestv = newv
qopt = qtest
return qopt
def target_function_gradient_numeric(self, x, loss=None, dx=1e-6):
if loss is None:
loss = self.loss
args_f = (self.edgels, self.i_param, loss)
return array([
(corisco_aux.angle_error(x + array([ dx,0,0,0]), *args_f) -
corisco_aux.angle_error(x + array([-dx,0,0,0]), *args_f)) / (2 * dx),
(corisco_aux.angle_error(x + array([0, dx,0,0]), *args_f) -
corisco_aux.angle_error(x + array([0,-dx,0,0]), *args_f)) / (2 * dx),
(corisco_aux.angle_error(x + array([0,0, dx,0]), *args_f) -
corisco_aux.angle_error(x + array([0,0,-dx,0]), *args_f)) / (2 * dx),
(corisco_aux.angle_error(x + array([0,0,0, dx]), *args_f) -
corisco_aux.angle_error(x + array([0,0,0,-dx]), *args_f)) / (2 * dx),
])
def ransac_search(self, initial_trials):
## Estimate solution using RANSAC
bestv = np.Inf ## Smallest value found
args_f = (self.edgels[~isnan(self.edgels[:,2])], self.i_param, self.loss)
for k in range(initial_trials):
## Pick indices of the reference normals. Re-sample until we
## get a list of three different values.
pk_a = np.random.random_integers(0,self.Ned-1)
pk_b = np.random.random_integers(0,self.Ned-1)
while pk_b == pk_a:
pk_b = np.random.random_integers(0,self.Ned-1)
pk_c = np.random.random_integers(0,self.Ned-1)
while pk_c == pk_a or pk_c == pk_b:
pk_c = np.random.random_integers(0,self.Ned-1)
## Get the normals with the first two chosen indices, and
## calculate a rotation matrix that has the x axis aligned to
## them.
n_a = self.normals[pk_a]
n_b = self.normals[pk_b]
vp1 = np.cross(n_a, n_b)
vp1 = vp1 * (vp1**2).sum()**-0.5
q1 = aligned_quaternion(vp1)
## Pick a new random third norm, and find the rotation to align
## the y direction to this edgel.
n_c = self.normals[pk_c]
vaux = np.dot(n_c, q1.rot())
ang = np.arctan2(vaux[1], -vaux[2])
q2 = Quat(np.sin(ang/2),0,0) * q1 ## The resulting orientation
## Find the value of the target function for this sampled
## orientation.
newv = corisco_aux.angle_error(q2.q, *args_f)
## If the value is the best yet, store solution.
if newv <= bestv :
bestv = newv
bpk_a = pk_a
bpk_b = pk_b
bpk_c = pk_c
qopt = q2
return qopt, bpk_a, bpk_b, bpk_c
def val_f(x,*fargs):
return corisco_aux.angle_error(x, fargs[0],fargs[1],fargs[2])
def target_function_value(self, x, loss=None):
if loss is None:
loss = self.loss
args_f = (self.edgels[~isnan(self.edgels[:,2])], self.i_param, loss)
# args_f = (self.edgels, self.i_param, loss)
return corisco_aux.angle_error(x, *args_f)
