def __call__(self,
beta,
theta,
exp=None,
reg_type='legacy',
get_skin=False,
name=None):
"""
Obtain SMPL with shape (beta) & pose (theta) inputs.
Theta includes the global rotation.
Args:
beta: N x 30
theta: N x 186 (with 3-D axis-angle rep)
exp:N x 200
Updates:
self.J_transformed: N x 24 x 3 joint location after shaping
& posing with beta and theta
Returns:
- joints: N x 19 or 14 x 3 joint locations depending on joint_type
If get_skin is True, also returns
- Verts: N x 6980 x 3
"""
num_batch = beta.shape[0]
# 1. Add shape blend shapes
# (N x 10) x (10 x 6890*3) = N x 6890 x 3
v_shaped = np.reshape(
np.matmul(beta, self.shapedirs),
[-1, self.size[0], self.size[1]]) + self.v_template
# 2. Infer shape-dependent joint locations.
Jx = np.matmul(v_shaped[:, :, 0], self.J_regressor)
Jy = np.matmul(v_shaped[:, :, 1], self.J_regressor)
Jz = np.matmul(v_shaped[:, :, 2], self.J_regressor)
J = np.stack([Jx, Jy, Jz], axis=2)
# 3. Add pose blend shapes
# N x 24 x 3 x 3
Rs = np.reshape(batch_rodrigues(np.reshape(theta, [-1, 3])),
[-1, self.nJoints, 3, 3])
# Ignore global rotation.
# No posedirs in Adam Model
# pose_feature = np.reshape(Rs[:, 1:, :, :] - np.eye(3),
# [-1, 207])
# # (N x 207) x (207, 20670) -> N x 6890 x 3
if exp is not None:
v_faced = np.reshape(np.matmul(exp, self.facedirs),
[-1, self.size[0], self.size[1]]) + v_shaped
else:
v_faced = v_shaped
# 4. Get the global joint location
self.J_transformed, A = batch_global_rigid_transformation(
Rs, J, self.parents)
# 5. Do skinning:
# W is N x 6890 x 24
W = np.reshape(np.tile(self.weights, [num_batch, 1]),
[num_batch, -1, self.nJoints])
# (N x 6890 x 24) x (N x 24 x 16)
T = np.reshape(
np.matmul(W, np.reshape(A, [num_batch, self.nJoints, 16])),
[num_batch, -1, 4, 4])
v_shaped_homo = np.concatenate(
[v_faced, np.ones([num_batch, v_faced.shape[1], 1])], 2)
v_homo = np.matmul(T, np.expand_dims(v_shaped_homo, -1))
verts = v_homo[:, :, :3, 0]
# Get cocoplus or lsp joints:
if reg_type == 'legacy':
joint_x = np.matmul(verts[:, :, 0], self.J_regressor)
joint_y = np.matmul(verts[:, :, 1], self.J_regressor)
joint_z = np.matmul(verts[:, :, 2], self.J_regressor)
joints = np.stack([joint_x, joint_y, joint_z], axis=2)
elif reg_type == 'coco25':
joint_x = np.matmul(verts[:, :, 0], self.J_regressor_coco25)
joint_y = np.matmul(verts[:, :, 1], self.J_regressor_coco25)
joint_z = np.matmul(verts[:, :, 2], self.J_regressor_coco25)
joints = np.stack([joint_x, joint_y, joint_z], axis=2)
else:
raise ValueError('Unknown regressor type')
if get_skin:
return verts, joints
else:
return joints
def call(self, theta, beta, trans, v_personal):
"""
Obtain SMPL with shape (beta) & pose (theta) inputs.
Theta includes the global rotation.
Args:
beta: N x 10
theta: N x 72 (with 3-D axis-angle rep)
Updates:
self.J_transformed: N x 24 x 3 joint location after shaping
& posing with beta and theta
Returns:
- joints: N x 19 or 14 x 3 joint locations depending on joint_type
If get_skin is True, also returns
- Verts: N x 6980 x 3 (low res) N x 27554 x 3 (hres)
"""
# Cast inputs to float64
theta, beta, trans, v_personal = tf.cast(
theta, self.data_type), tf.cast(beta, self.data_type), tf.cast(
trans, self.data_type), tf.cast(v_personal, self.data_type)
if not self.isHres:
v_personal = v_personal[:, :6890, :]
# num_batch = int(tf.shape(beta)[0])
num_batch = tf.keras.backend.int_shape(beta)[0]
# 1. Add shape blend shapes
v_shaped_scaled = tf.reshape(
tf.matmul(beta, self.shapedirs, name='shape_bs'),
[-1, self.size[0], self.size[1]]) + self.v_template
if self.scale:
body_height = (
v_shaped_scaled[:, 2802, 1] + v_shaped_scaled[:, 6262, 1]) - (
v_shaped_scaled[:, 2237, 1] + v_shaped_scaled[:, 6728, 1])
scale = tf.reshape(1.66 / body_height, (-1, 1, 1))
else:
scale = tf.reshape(tf.ones_like(v_shaped_scaled[:, 2802, 1]),
(-1, 1, 1))
# Scale to 1.66m height
v_shaped_scaled *= scale
v_shaped = v_shaped_scaled
if self.isHres:
v_shaped = tf.map_fn(self.fn, v_shaped, dtype=self.data_type)
v_shaped_personal = v_shaped + v_personal
# 2. Infer shape-dependent joint locations.
# Some gpu dont support float64 operations
with tf.device('/cpu:0'):
Jx = tf.transpose(
tf.sparse_tensor_dense_matmul(
self.J_regressor, tf.transpose(v_shaped_scaled[:, :, 0])))
Jy = tf.transpose(
tf.sparse_tensor_dense_matmul(
self.J_regressor, tf.transpose(v_shaped_scaled[:, :, 1])))
Jz = tf.transpose(
tf.sparse_tensor_dense_matmul(
self.J_regressor, tf.transpose(v_shaped_scaled[:, :, 2])))
J = tf.stack([Jx, Jy, Jz], axis=2)
# 3. Add pose blend shapes
# N x 24 x 3 x 3
if self.theta_in_rodrigues:
Rs = tf.reshape(batch_rodrigues(tf.reshape(theta, [-1, 3])),
[-1, 24, 3, 3])
else:
if self.theta_is_perfect_rotmtx:
Rs = theta
else:
s, u, v = tf.svd(theta)
Rs = tf.matmul(u, tf.transpose(v, perm=[0, 1, 3, 2]))
# with tf.name_scope("lrotmin"):
# Ignore global rotation.
pose_feature = tf.reshape(Rs[:, 1:, :, :] - tf.eye(3, dtype=Rs.dtype),
[-1, 207])
# (N x 207) x (207, 20670) -> N x 6890 x 3
v_posed = tf.reshape(tf.matmul(pose_feature, self.posedirs),
[-1, self.size[0], self.size[1]])
if self.isHres:
v_posed = tf.map_fn(self.fn, v_posed)
v_posed += v_shaped_personal
# 4. Get the global joint location
J_transformed, A = batch_global_rigid_transformation(
Rs, J, self.parents)
J_transformed += tf.expand_dims(trans, axis=1)
# 5. Do skinning:
if self.isHres:
W = tf.reshape(tf.tile(self.weights_hres, [num_batch, 1]),
[num_batch, -1, 24])
else:
# W is N x 6890 x 24
W = tf.reshape(tf.tile(self.weights_, [num_batch, 1]),
[num_batch, -1, 24])
# (N x 6890 x 24) x (N x 24 x 16)
T = tf.reshape(tf.matmul(W, tf.reshape(A, [num_batch, 24, 16])),
[num_batch, -1, 4, 4])
v_posed_homo = tf.concat([
v_posed,
tf.ones([num_batch, v_posed.shape[1], 1], dtype=v_posed.dtype)
], 2)
v_homo = tf.matmul(T, tf.expand_dims(v_posed_homo, -1))
verts = v_homo[:, :, :3, 0] # / v_homo[:, :, 3, 0:1]
verts_t = verts + tf.expand_dims(trans, axis=1)
# Return verts, unposed verts, unposed naked verts, joints
return verts_t, v_shaped_personal, v_shaped, J_transformed
def __call__(self, beta, theta, get_skin=False, name=None):
"""
Obtain SMPL with shape (beta) & pose (theta) inputs.
Theta includes the global rotation.
Args:
beta: N x 10
theta: N x 72 (with 3-D axis-angle rep)
Updates:
self.J_transformed: N x 24 x 3 joint location after shaping
& posing with beta and theta
Returns:
- joints: N x 19 or 14 x 3 joint locations depending on joint_type
If get_skin is True, also returns
- Verts: N x 6980 x 3
"""
with tf.compat.v1.variable_scope(name, "smpl_main", [beta, theta]):
num_batch = tf.shape(input=beta)[0]
#print('\n\n\n\n','Num batch = ',num_batch)
#print(beta.shape, tf.shape(beta)[1], beta.shape[0], beta.shape[1])
# 1. Add shape blend shapes
# (N x 10) x (10 x 6890*3) = N x 6890 x 3
v_shaped = tf.reshape(
tf.matmul(beta, self.shapedirs, name='shape_bs'),
[-1, self.size[0], self.size[1]]) + self.v_template
# 2. Infer shape-dependent joint locations.
Jx = tf.matmul(v_shaped[:, :, 0], self.J_regressor)
Jy = tf.matmul(v_shaped[:, :, 1], self.J_regressor)
Jz = tf.matmul(v_shaped[:, :, 2], self.J_regressor)
J = tf.stack([Jx, Jy, Jz], axis=2)
# 3. Add pose blend shapes
# N x 24 x 3 x 3
Rs = tf.reshape(batch_rodrigues(tf.reshape(theta, [-1, 3])),
[-1, 24, 3, 3])
with tf.compat.v1.variable_scope("lrotmin"):
# Ignore global rotation.
pose_feature = tf.reshape(Rs[:, 1:, :, :] - tf.eye(3),
[-1, 207])
# (N x 207) x (207, 20670) -> N x 6890 x 3
v_posed = tf.reshape(tf.matmul(pose_feature, self.posedirs),
[-1, self.size[0], self.size[1]]) + v_shaped
#4. Get the global joint location
self.J_transformed, A = batch_global_rigid_transformation(
Rs, J, self.parents)
# 5. Do skinning:
# W is N x 6890 x 24, the weight of 24 joints inference every vertex
W = tf.reshape(tf.tile(self.weights1, [num_batch, 1]),
[num_batch, -1, 24])
#print(W[0,0:2])
# (N x 6890 x 24) x (N x 24 x 16)
T = tf.reshape(tf.matmul(W, tf.reshape(A, [num_batch, 24, 16])),
[num_batch, -1, 4, 4])
v_posed_homo = tf.concat(
[v_posed, tf.ones([num_batch, v_posed.shape[1], 1])], 2)
v_homo = tf.matmul(T, tf.expand_dims(v_posed_homo, -1))
#print(v_homo)
verts = v_homo[:, :, :3, 0]
# Get cocoplus or lsp joints:
joint_x = tf.matmul(verts[:, :, 0], self.joint_regressor)
joint_y = tf.matmul(verts[:, :, 1], self.joint_regressor)
joint_z = tf.matmul(verts[:, :, 2], self.joint_regressor)
joints = tf.stack([joint_x, joint_y, joint_z], axis=2)
#print(verts.shape,'\n\n\n')
if get_skin:
return verts, joints, Rs
else:
return joints