def _computeInitialNormal(self):
A_8U = self._A_8U
h, w = A_8U.shape
N0_32F = silhouetteNormal(A_8U)
A, b = normalConstraints(A_8U, N0_32F)
N_flat = amg_solver.solve(A, b)
N_flat = normalizeVectors(N_flat)
N_32F = N_flat.reshape(h, w, 3)
self._N0_32F = N_32F
def _computeInitialNormal(self):
A_8U = self._A_8U
h, w = A_8U.shape
N0_32F = silhouetteNormal(A_8U)
A, b = normalConstraints(A_8U, N0_32F)
N_flat = amg_solver.solve(A, b)
N_flat = normalizeVectors(N_flat)
N_32F = N_flat.reshape(h, w, 3)
self._N0_32F = N_32F
def estimateNormal(A_8U):
h, w = A_8U.shape
N0_32F = silhouetteNormal(A_8U)
A, b = normalConstraints(A_8U, N0_32F)
N_flat = amg_solver.solve(A, b)
N_flat = computeNz(N_flat)
N_flat = normalizeVectors(N_flat)
N_32F = N_flat.reshape(h, w, 3)
return N0_32F, N_32F
def _optimize(self, A_I, b_I):
h, w = self._I_32F.shape[:2]
A_L = laplacianMatrix((h, w), num_elements=1)
A_L = A_L
A = A_L + A_I
b = b_I
N = amg_solver.solve(A, b).reshape(-1, 3)
computeNz(N)
#N = normalizeVectors(N)
N_32F = N.reshape(h, w, 3)
self._N_32F = N_32F
def _interpolateNormalAMG(self, N0_32F, W_32F, A_8U):
h, w = N0_32F.shape[:2]
A_c, b_c = normalConstraints(W_32F, N0_32F)
A_8U = None
if self._image.shape[2] == 4:
A_8U = to8U(alpha(self._image))
A_sil, b_sil = silhouetteConstraints(A_8U)
A_L = laplacianMatrix((h, w))
A = 10.0 * A_c + A_L + A_sil
b = 10.0 * b_c + b_sil
N_32F = amg_solver.solve(A, b).reshape(h, w, 3)
N_32F = normalizeImage(N_32F)
return N_32F
def depthFromGradient(I_32F, A_8U):
if A_8U is not None:
I_32F = preProcess(I_32F, A_8U)
h, w = I_32F.shape[:2]
A_g, b_g = gradientConstraints(I_32F, w_cons=1.0)
A_rg = l2Regularization(h*w, w_rg=0.000001)
A_L = laplacianMatrix((h, w))
A = A_g + A_rg
b = b_g
D_flat = amg_solver.solve(A, b)
D_32F = D_flat.reshape(h, w)
return D_32F
def _computeLumoNormal(self):
A_8U = self._A_8U
if A_8U is None:
return
h, w = A_8U.shape[:2]
A_c, b_c = amg_constraints.silhouetteConstraints(A_8U)
A_L = amg_constraints.laplacianMatrix((h, w))
A = 3.0 * A_c + A_L
b = 3.0 * b_c
N_32F = amg_solver.solve(A, b).reshape(h, w, 3)
N_32F = computeNz(N_32F.reshape(-1, 3)).reshape(h, w, 3)
N_32F = normalizeImage(N_32F)
self._N_lumo = np.array(N_32F)
def depthFromGradient(I_32F, A_8U):
if A_8U is not None:
I_32F = preProcess(I_32F, A_8U)
h, w = I_32F.shape[:2]
A_g, b_g = gradientConstraints(I_32F, w_cons=1.0)
A_rg = l2Regularization(h * w, w_rg=0.000001)
A_L = laplacianMatrix((h, w))
A = A_g + A_rg
b = b_g
D_flat = amg_solver.solve(A, b)
D_32F = D_flat.reshape(h, w)
return D_32F
def _computeLumoNormal(self):
A_8U = self._A_8U
if A_8U is None:
return
h, w = A_8U.shape[:2]
A_c, b_c = amg_constraints.silhouetteConstraints(A_8U)
A_L = amg_constraints.laplacianMatrix((h, w))
A = 3.0 * A_c + A_L
b = 3.0 * b_c
N_32F = amg_solver.solve(A, b).reshape(h, w, 3)
N_32F = computeNz(N_32F.reshape(-1, 3)).reshape(h, w, 3)
N_32F = normalizeImage(N_32F)
self._N_lumo = np.array(N_32F)
def _computeInitialNormal(self):
A_8U = self._A_8U
self._N0_32F = np.float64(silhouetteNormal(A_8U))
return
h, w = A_8U.shape
A_L = laplacianMatrix((h, w), num_elements=3)
A_sil, b_sil = silhouetteConstraints(A_8U, is_flat=True)
A = A_L + A_sil
b = b_sil
N = amg_solver.solve(A, b).reshape(-1, 3)
computeNz(N)
N = normalizeVectors(N)
N_32F = N.reshape(h, w, 3)
self._N0_32F = N_32F
def _computeInitialNormal(self):
A_8U = self._A_8U
self._N0_32F = np.float64(silhouetteNormal(A_8U))
return
h, w = A_8U.shape
A_L = laplacianMatrix((h, w), num_elements=3)
A_sil, b_sil = silhouetteConstraints(A_8U, is_flat=True)
A = A_L + A_sil
b = b_sil
N = amg_solver.solve(A, b).reshape(-1, 3)
computeNz(N)
N = normalizeVectors(N)
N_32F = N.reshape(h, w, 3)
self._N0_32F = N_32F
def smoothGradient(gx):
h, w = gx.shape[:2]
A_L = laplacianMatrix((h, w))
w_g = 1.0
gx_flat = gx.flatten()
cons_ids = np.where(gx_flat > 0.01 * np.max(0.0))
num_verts = h * w
diags = np.zeros(num_verts)
diags[cons_ids] = w_g
A = scipy.sparse.diags(diags, 0) + A_L
b = np.zeros(num_verts, dtype=np.float32)
b[cons_ids] = w_g * gx_flat[cons_ids]
gx_flat = amg_solver.solve(A, b)
gx = gx_flat.reshape(h, w)
print gx.shape
return np.float32(gx)
def smoothGradient(gx):
h, w = gx.shape[:2]
A_L = laplacianMatrix((h, w))
w_g = 1.0
gx_flat = gx.flatten()
cons_ids = np.where(gx_flat > 0.01 * np.max(0.0))
num_verts = h * w
diags = np.zeros(num_verts)
diags[cons_ids] = w_g
A = scipy.sparse.diags(diags, 0) + A_L
b = np.zeros(num_verts, dtype=np.float32)
b[cons_ids] = w_g * gx_flat[cons_ids]
gx_flat = amg_solver.solve(A, b)
gx = gx_flat.reshape(h, w)
print gx.shape
return np.float32(gx)
def _runImp(self):
normal_data = loadNormal(self._data_file)
if normal_data is None:
return
N0_32F, A_8U = normal_data
h, w = A_8U.shape[:2]
A_c, b_c = amg_constraints.silhouetteConstraints(A_8U, is_flat=True)
A_L = amg_constraints.laplacianMatrix((h, w), num_elements=3)
A = A_c + A_L
b = b_c
N = amg_solver.solve(A, b).reshape(-1, 3)
N = computeNz(N)
N = normalizeVectors(N)
N_32F = N.reshape(h, w, 3)
file_path = self.resultFile(self._data_file_name)
saveNormal(file_path, N_32F, A_8U)
def _runImp(self):
normal_data = loadNormal(self._data_file)
if normal_data is None:
return
N0_32F, A_8U = normal_data
h, w = A_8U.shape[:2]
A_c, b_c = amg_constraints.silhouetteConstraints(A_8U, is_flat=True)
A_L = amg_constraints.laplacianMatrix((h, w), num_elements=3)
A = A_c + A_L
b = b_c
N = amg_solver.solve(A, b).reshape(-1, 3)
N = computeNz(N)
N = normalizeVectors(N)
N_32F = N.reshape(h, w, 3)
file_path = self.resultFile(self._data_file_name)
saveNormal(file_path, N_32F, A_8U)
def _computeDetailNormal(self, N0_32F):
h, w = N0_32F.shape[:2]
W_32F = np.zeros((h, w))
# sigma_d = 2.0 * np.max(N0_32F[:, :, 2])
# W_32F = 1.0 - np.exp( - (N0_32F[:, :, 2] ** 2) / (sigma_d ** 2))
W_32F = 1.0 - N0_32F[:, :, 2]
W_32F *= 1.0 / np.max(W_32F)
W_32F = W_32F ** 1.5
A_c, b_c = amg_constraints.normalConstraints(W_32F, N0_32F)
A_L = amg_constraints.laplacianMatrix((h, w))
lambda_d = 2.0
A = A_c + lambda_d * A_L
b = b_c
N_32F = amg_solver.solve(A, b).reshape(h, w, 3)
N_32F = computeNz(N_32F.reshape(-1, 3)).reshape(h, w, 3)
N_32F = normalizeImage(N_32F)
return N_32F
def _computeDetailNormal(self, N0_32F):
h, w = N0_32F.shape[:2]
W_32F = np.zeros((h, w))
# sigma_d = 2.0 * np.max(N0_32F[:, :, 2])
# W_32F = 1.0 - np.exp( - (N0_32F[:, :, 2] ** 2) / (sigma_d ** 2))
W_32F = 1.0 - N0_32F[:, :, 2]
W_32F *= 1.0 / np.max(W_32F)
W_32F = W_32F**1.5
A_c, b_c = amg_constraints.normalConstraints(W_32F, N0_32F)
A_L = amg_constraints.laplacianMatrix((h, w))
lambda_d = 2.0
A = A_c + lambda_d * A_L
b = b_c
N_32F = amg_solver.solve(A, b).reshape(h, w, 3)
N_32F = computeNz(N_32F.reshape(-1, 3)).reshape(h, w, 3)
N_32F = normalizeImage(N_32F)
return N_32F
