def train(cls, lmks, imgs, ref, frac, kmax):
smodel = ShapeModel.load('data/shape.npz')
tmodel = TextureModel.load('data/texture.npz')
# build diagonal matrix of weights that measures
# the unit difference between shape and texture parameters
r = tmodel.variance.sum() / smodel.variance.sum()
Ws = r * np.eye(smodel.num_modes())
n_samples = lmks.shape[1]
tm = TextureMapper(480, 640)
# create empty matrix
n_feats = smodel.num_modes() + tmodel.num_modes()
A = np.empty((n_feats, n_samples))
# concatenate shape and texture parameters for each training example
for i in xrange(n_samples):
img = next(imgs)
lmk = lmks[:,i]
sparams = smodel.calc_params(lmk)
tparams = tmodel.calc_params(img, lmk, ref, tm.warp_triangles)
# ignore first 4 shape parameters
A[:,i] = np.concatenate((Ws.dot(sparams[4:]), tparams))
D = pca(A, frac, kmax)
# compute variance
Q = pow(D.T.dot(A), 2)
e = Q.sum(axis=1) / (n_samples-1)
return cls(D, e, Ws)
def view_combined_model(shp_fn, txt_fn, cmb_fn, scale, tranx, trany, width, height):
img = np.empty((height, width, 3), dtype='uint8')
# cv2.namedWindow('combined model')
tm = TextureMapper(img.shape[1], img.shape[0])
smodel = ShapeModel.load(shp_fn)
tmodel = TextureModel.load(txt_fn)
cmodel = CombinedModel.load(cmb_fn)
vals = genvals()
params = smodel.get_params(scale, tranx, trany)
ref = smodel.calc_shape(params)
ref = ref.reshape((ref.size // 2, 2))
verts = get_vertices(ref)
while True:
for k in range(cmodel.num_modes()):
for v in vals:
p = np.zeros(cmodel.num_modes())
p[k] = v * 3 * np.sqrt(cmodel.variance[k])
sparams, tparams = cmodel.calc_shp_tex_params(p, smodel.num_modes())
params[4:] = sparams
shp = smodel.calc_shape(params)
shp = shp.reshape(ref.shape)
t = tmodel.calc_texture(tparams)
img[:] = 0
draw_texture(img, t, ref)
warped = tm.warp_triangles(img, ref[verts], shp[verts])
s = 'mode: %d, val: %f sd' % (k, v * 3)
draw_string(warped, s)
print('#####')
# cv2.imshow('combined model', warped)
cv2.imwrite('output/%d.jpg'%(int(time.time()*1000000)), warped)
def view_combined_model(shp_fn, txt_fn, cmb_fn, scale, tranx, trany, width, height):
img = np.empty((height, width, 3), dtype='uint8')
cv2.namedWindow('combined model')
tm = TextureMapper(img.shape[1], img.shape[0])
smodel = ShapeModel.load(shp_fn)
tmodel = TextureModel.load(txt_fn)
cmodel = CombinedModel.load(cmb_fn)
vals = genvals()
params = smodel.get_params(scale, tranx, trany)
ref = smodel.calc_shape(params)
ref = ref.reshape((ref.size//2, 2))
verts = get_vertices(ref)
while True:
for k in xrange(cmodel.num_modes()):
for v in vals:
p = np.zeros(cmodel.num_modes())
p[k] = v * 3 * np.sqrt(cmodel.variance[k])
sparams, tparams = cmodel.calc_shp_tex_params(p, smodel.num_modes())
params[4:] = sparams
shp = smodel.calc_shape(params)
shp = shp.reshape(ref.shape)
t = tmodel.calc_texture(tparams)
img[:] = 0
draw_texture(img, t, ref)
warped = tm.warp_triangles(img, ref[verts], shp[verts])
s = 'mode: %d, val: %f sd' % (k, v*3)
draw_string(warped, s)
cv2.imshow('combined model', warped)
if cv2.waitKey(10) == 27:
sys.exit()
def train(cls, lmks, imgs, ref, frac, kmax):
smodel = ShapeModel.load('data/shape.npz')
tmodel = TextureModel.load('data/texture.npz')
# build diagonal matrix of weights that measures
# the unit difference between shape and texture parameters
r = tmodel.variance.sum() / smodel.variance.sum()
Ws = r * np.eye(smodel.num_modes())
n_samples = lmks.shape[1]
tm = TextureMapper(480, 640)
# create empty matrix
n_feats = smodel.num_modes() + tmodel.num_modes()
A = np.empty((n_feats, n_samples))
# concatenate shape and texture parameters for each training example
for i in xrange(n_samples):
img = next(imgs)
lmk = lmks[:, i]
sparams = smodel.calc_params(lmk)
tparams = tmodel.calc_params(img, lmk, ref, tm.warp_triangles)
# ignore first 4 shape parameters
A[:, i] = np.concatenate((Ws.dot(sparams[4:]), tparams))
D = pca(A, frac, kmax)
# compute variance
Q = pow(D.T.dot(A), 2)
e = Q.sum(axis=1) / (n_samples - 1)
return cls(D, e, Ws)
def view_texture_model(shp_fn, txt_fn, scale, tranx, trany, width, height):
img = np.empty((height, width, 3), dtype='uint8')
smodel = ShapeModel.load(shp_fn)
tmodel = TextureModel.load(txt_fn)
vals = genvals()
# get reference shape
ref = smodel.get_shape(scale, tranx, trany)
ref = ref.reshape((ref.size // 2, 2))
while True:
for k in range(tmodel.num_modes()):
for v in vals:
p = np.zeros(tmodel.num_modes())
p[k] = v * 3 * np.sqrt(tmodel.variance[k])
img[:] = 0
s = 'mode: %d, val: %f sd' % (k, v * 3)
draw_string(img, s)
t = tmodel.calc_texture(p)
draw_texture(img, t, ref)
cv2.imshow('texture model', img)
if cv2.waitKey(10) == 27:
sys.exit()
def view_texture_model(shp_fn, txt_fn, scale, tranx, trany, width, height):
img = np.empty((height, width, 3), dtype='uint8')
smodel = ShapeModel.load(shp_fn)
tmodel = TextureModel.load(txt_fn)
vals = genvals()
# get reference shape
ref = smodel.get_shape(scale, tranx, trany)
ref = ref.reshape((ref.size//2, 2))
while True:
for k in xrange(tmodel.num_modes()):
for v in vals:
p = np.zeros(tmodel.num_modes())
p[k] = v * 3 * np.sqrt(tmodel.variance[k])
img[:] = 0
s = 'mode: %d, val: %f sd' % (k, v*3)
draw_string(img, s)
t = tmodel.calc_texture(p)
draw_texture(img, t, ref)
cv2.imshow('texture model', img)
if cv2.waitKey(10) == 27:
sys.exit()
help='output file')
parser.add_argument('--no-flip', action='store_false', dest='flipped',
help='exclude flipped data')
return parser.parse_args()
if __name__ == '__main__':
args = parse_args()
muct = MuctDataset()
muct.load(clean=True)
# If a face is too close to the image border and we perturbe
# the scale then that face may grow beyond the image border.
# Ignore problematic images.
muct.ignore('i405wc-fn')
data = muct.all_lmks()
imgs = muct.iterimages(mirror=True)
print 'training samples:', len(data)
smodel = ShapeModel.load('data/shape.npz')
tmodel = TextureModel.load('data/texture.npz')
# get reference shape
params = smodel.get_params(200, 150, 150)
ref = smodel.calc_shape(params)
ref = ref.reshape((ref.size//2, 2))
experiments(imgs, data, smodel, tmodel, ref, args.fout)
print 'wrote', args.fout