def fit_submanifold_landmarks_to_image(template,
original,
Xlm,
face_d,
face_p,
landmarks=list(range(68))):
'''
Fit the submanifold to the template and take the top-K.
Xlm is a N x 68 x 2 list of landmarks.
'''
lossX = numpy.empty((len(Xlm), ), dtype=numpy.float64)
MX = numpy.empty((len(Xlm), 2, 3), dtype=numpy.float64)
nfail = 0
for i in range(len(Xlm)):
lm = Xlm[i]
try:
M, loss = alignface.fit_face_landmarks(
Xlm[i],
template,
landmarks=landmarks,
image_dims=original.shape[:2])
lossX[i] = loss
MX[i] = M
except alignface.FitError:
lossX[i] = float('inf')
MX[i] = 0
nfail += 1
if nfail > 1:
print('fit submanifold, {} errors.'.format(nfail))
a = numpy.argsort(lossX)
return a, lossX, MX
responseid = args[1]
args = args[2:]
if command == 'ping':
channel.send((('okay', responseid), {}))
elif command == 'forward_landmarks':
# X is pickled N x H x W x 3 images
# Xlm is pickled N x 68 x 2 landmarks
# scores is pickled N x K attribute scores
X = pickle.loads(kwargs['X'])
Xlm = pickle.loads(kwargs['Xlm'])
result = []
for x, lm in zip(X, Xlm):
# lm,x=alignface.detect_landmarks('../../images/jpg-816x1000/vvdd-11.jpg',face_d,face_p)
M, loss = alignface.fit_face_landmarks(
lm,
fm.celeba_template,
image_dims=[160, 160],
twoscale=False)
Y = alignface.warp_to_template(x, M, image_dims=[160, 160])
Y = numpy.expand_dims(Y, 0)
Y = fm.preprocess(Y)
Y = fm.forward(Y)
scores = fm.predict_attributes(Y)
result.append(scores[0])
result = numpy.asarray(result)
channel.send((('okay', responseid), {
'scores': pickle.dumps(result)
}))
elif command == 'forward_images':
# X is pickled list of N strings
# scores is pickled N x K attributes scores