def predict(self, img, pre_label, show=False):
state = [copy.deepcopy(self.f_value[0][0])]
if show:
showpic(img, label_to_landmark(state))
phi_star = self.get_feature(img, pre_label)
for i in range(len(self.derection)):
deta_X = pre_label - state
deta_phi = phi_star - self.get_feature(img, state)
pca_deta_X = np.dot(self.pca_base[i][0], np.transpose(deta_X))
print(pca_deta_X.shape)
pca_deta_phi = np.dot(self.pca_base[i][1], np.transpose(deta_phi))
if pca_deta_X[0] >= 0:
if pca_deta_X[1] >= 0:
if pca_deta_phi[0] >= 0:
subindex = 0
else:
subindex = 1
else:
if pca_deta_phi[0] >= 0:
subindex = 2
else:
subindex = 3
else:
if pca_deta_X[1] >= 0:
if pca_deta_phi[0] >= 0:
subindex = 4
else:
subindex = 5
else:
if pca_deta_phi[0] >= 0:
subindex = 6
else:
subindex = 7
state += self.learning_rate * self.compute_strid(
img, state, self.derection[i][subindex])
if show:
showpic(img, label_to_landmark(state))
return state
def compute_strid(self,img, state, decent_map):
'''
输入单张图片 和 标注 利用计算出来的每一步方向求出每一步的步伐
:param imgs:
:param state:
:param decent_map:
:return:
'''
phi = get_hog_feature(img, label_to_landmark(state),self.feature_block) # (4-1)*(4-1)*9*4*12 =3888
phi = np.array(phi)
phi_one = np.ones(phi.shape)
A = np.hstack((phi, phi_one))
A = np.mat(A)
return np.dot(A, decent_map)
def compute_strid(self,img, state, decent_map):
'''
输入单张图片 和 标注 利用计算出来的每一步方向求出每一步的步伐
:param imgs:
:param state:
:param decent_map:
:return:
'''
R = decent_map
phi = get_hog_feature(img, label_to_landmark(state)) # (4-1)*(4-1)*9*4*12 =3888
phi = np.array(phi)
A = phi-self.avg_phi_star
return np.dot(A, R)
def compute_strid_batch(self,imgs, states, decent_map):
'''
输入多张图片 和 标注 利用计算出来的每一步方向求出每一步的步伐
:param imgs:
:param state:
:param decent_map:
:return:
'''
R = decent_map
phi = [get_hog_feature(imgs[idx], label_to_landmark(states[idx]))for idx in range(len(imgs))] # (4-1)*(4-1)*9*4*12 =3888
phi = np.array(phi)
train_phi_star = [self.avg_phi_star for idx in range(len(imgs))]
train_phi_star = np.array(train_phi_star)
A = phi-train_phi_star
return np.dot(A, R)
def get_phis(self,imgs,label):
phi_star = [get_hog_feature(imgs[idx], label_to_landmark(label[idx]),self.feature_block) for idx in
range(len(imgs))] # (4-1)*(4-1)*9*4*12 =3888
phi_star = np.array(phi_star)
return phi_star