def gen_meanpose(phase, config, n_samp=20000):
data_dir = config.train_dir if phase == "train" else config.test_dir
all_paths = glob.glob(os.path.join(data_dir, '*/*/motions/*.npy'))
random.shuffle(all_paths)
all_paths = all_paths[:n_samp]
all_joints = []
print("computing meanpose and stdpose")
for path in tqdm(all_paths):
try:
motion = np.load(path)
except:
continue
if motion.shape[1] == 3:
basis = None
if sum(config.rotation_axes) > 0:
x_angles = view_angles if config.rotation_axes[
0] else np.array([0])
z_angles = view_angles if config.rotation_axes[
1] else np.array([0])
y_angles = view_angles if config.rotation_axes[
2] else np.array([0])
x_angles, z_angles, y_angles = np.meshgrid(
x_angles, z_angles, y_angles)
angles = np.stack([
x_angles.flatten(),
z_angles.flatten(),
y_angles.flatten()
],
axis=1)
i = np.random.choice(len(angles))
basis = get_change_of_basis(motion, angles[i])
motion = preprocess_mixamo(motion)
motion = rotate_motion_3d(motion, basis)
motion = localize_motion(motion)
all_joints.append(motion)
else:
motion = preprocess_mixamo(motion)
motion = rotate_motion_3d(motion, basis)
motion = localize_motion(motion)
all_joints.append(motion)
else:
motion = motion * 128
motion_proj = localize_motion(motion)
all_joints.append(motion_proj)
all_joints = np.concatenate(all_joints, axis=2)
meanpose = np.mean(all_joints, axis=2)
stdpose = np.std(all_joints, axis=2)
stdpose[np.where(stdpose == 0)] = 1e-9
return meanpose, stdpose
def preprocessing(self, motion3d, view_angle=None, params=None):
if self.aug: motion3d, params = self.augmentation(motion3d, params)
basis = None
if view_angle is not None:
basis = get_change_of_basis(motion3d, view_angle)
motion3d = preprocess_mixamo(motion3d)
motion3d = rotate_motion_3d(motion3d, basis)
motion2d = motion3d[:, [0, 2], :]
motion2d_scale = limb_scale_motion_2d(motion2d, self.global_range,
self.local_range)
motion2d = localize_motion(motion2d)
motion2d_scale = localize_motion(motion2d_scale)
motion2d = normalize_motion(motion2d, self.meanpose, self.stdpose)
motion2d_scale = normalize_motion(motion2d_scale, self.meanpose,
self.stdpose)
motion2d = motion2d.reshape([-1, motion2d.shape[-1]])
motion2d_scale = motion2d_scale.reshape((-1, motion2d_scale.shape[-1]))
motion2d = torch.from_numpy(motion2d).float()
motion2d_scale = torch.from_numpy(motion2d_scale).float()
return motion2d, motion2d_scale
def preprocessing(self, motion3d, view_angle=None, params=None):
"""
:param item: filename built from self.build_tiem
:return:
"""
if self.aug: motion3d, params = self.augmentation(motion3d, params)
basis = None
if view_angle is not None:
basis = get_change_of_basis(motion3d, view_angle)
motion3d = preprocess_mixamo(motion3d)
motion3d = rotate_motion_3d(motion3d, basis)
motion3d = localize_motion(motion3d)
motion3d = normalize_motion(motion3d, self.meanpose, self.stdpose)
motion2d = motion3d[:, [0, 2], :]
motion3d = motion3d.reshape([-1, motion3d.shape[-1]])
motion2d = motion2d.reshape([-1, motion2d.shape[-1]])
motion3d = torch.from_numpy(motion3d).float()
motion2d = torch.from_numpy(motion2d).float()
return motion3d, motion2d
def load_and_preprocess(path):
motion3d = np.load(path)
# length must be multiples of 8 due to the size of convolution
_, _, T = motion3d.shape
T = (T // 8) * 8
motion3d = motion3d[:, :, :T]
# project to 2d
motion_proj = motion3d[:, [0, 2], :]
# reformat for mixamo data
motion_proj = preprocess_mixamo(motion_proj, unit=1.0)
return motion_proj
def load_and_preprocess(path, config, mean_pose, std_pose):
motion3d = np.load(path)
# length must be multiples of 8 due to the size of convolution
_, _, T = motion3d.shape
T = (T // 8) * 8
motion3d = motion3d[:, :, :T]
# project to 2d
motion_proj = motion3d[:, [0, 2], :]
# reformat for mixamo data
motion_proj = preprocess_mixamo(motion_proj, unit=1.0)
# preprocess for network input
motion_proj, start = preprocess_test(motion_proj, mean_pose, std_pose,
config.data.unit)
motion_proj = motion_proj.reshape((-1, motion_proj.shape[-1]))
motion_proj = torch.from_numpy(motion_proj).float()
return motion_proj, start