File size: 8,927 Bytes
26225c5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 |
import torch
from src.data import NAG
from src.transforms import Transform
from src.utils.geometry import rodrigues_rotation_matrix
__all__ = [
'CenterPosition', 'RandomTiltAndRotate', 'RandomAnisotropicScale',
'RandomAxisFlip']
class CenterPosition(Transform):
"""Center the position of all nodes of all levels of a NAG around
their level-0 centroid.
"""
_IN_TYPE = NAG
_OUT_TYPE = NAG
def _process(self, nag):
offset = nag[0].pos.mean(dim=0)
for i_level in range(nag.num_levels):
nag[i_level].pos -= offset
return nag
class RandomTiltAndRotate(Transform):
"""Rotate the NAG around a random axis, with a random angle. The
axis is picked following a gaussian jitter around the z axis. The
angle is picked following a uniform distribution within a specified
range.
If the nodes have a `normal` or 'mean_normal' attribute, we also
rotate those accordingly.
Warning: any other absolute orientation-related attributes beside
`pos`, `normal` and 'mean_normal' may be broken by this transform.
:param phi: float (degrees)
The random axis will have random angle wrt the z axis. This
random angle corresponds to adding some random xy offset to z.
This offset is sampled from a 2D gaussian distribution of
standard deviation `sigma` computed so that a `3 * sigma` xy
offset corresponds to a `phi` angle.
:param theta: float (degrees)
The random rotation angle will be uniformly picked within
[-abs(theta), abs(theta)]
"""
_IN_TYPE = NAG
_OUT_TYPE = NAG
def __init__(self, phi=5, theta=180):
assert isinstance(phi, (int, float))
assert isinstance(theta, (int, float))
self.phi = float(abs(phi))
self.theta = float(abs(theta))
def _process(self, nag):
device = nag.device
# Generate the random rotation axis
sigma = self.phi / 180. * torch.pi / 3
if sigma > 0:
means = torch.zeros(2, device=device)
stds = torch.eye(2, device=device) * sigma
distribution = torch.distributions.MultivariateNormal(means, stds)
axis_xy = distribution.sample()
axis_z = torch.ones(1, device=device)
axis = torch.cat((axis_xy, axis_z))
axis /= axis.norm()
else:
axis = torch.zeros(3, device=device, dtype=torch.float)
axis[2] = 1
# Generate the random rotation angle
theta = torch.rand(1, device=device) * 2 * self.theta - self.theta
# Compute the rotation matrix
R = rodrigues_rotation_matrix(axis, theta)
# Rotate the nodes at each level. If the nodes have a `normal`
# attribute, we also rotate those accordingly
for i_level in range(nag.num_levels):
if sigma <= 0:
continue
nag[i_level].pos = nag[i_level].pos @ R.T
# If the nodes have a `normal` or 'mean_normal' attribute,
# we also adapt their orientations accordingly
for k in ['normal', 'mean_normal']:
if getattr(nag[i_level], k, None) is not None:
nag[i_level][k] = self._rotate_normal(nag[i_level][k], R)
# TODO: this is an ugly, hardcoded patch to deal with
# features assumedly created by
# _minimalistic_horizontal_edge_features........
if nag[i_level].edge_attr is not None:
edge_attr = nag[i_level].edge_attr
assert edge_attr.shape[1] == 7, \
"Expected exactly 7 features in `edge_attr`, generated " \
"with `_minimalistic_horizontal_edge_features`"
dtype = edge_attr.dtype
edge_attr[:, :3] = (edge_attr[:, :3].float() @ R.T).to(dtype) # `mean_off`, float16 mm not supported on CPU
nag[i_level].edge_attr = edge_attr
return nag
@staticmethod
def _rotate_normal(normal, R):
dtype = normal.dtype
normal = (normal.float() @ R.T).to(dtype)
normal[normal[:, 2] < 0] *= -1
return normal
class RandomAnisotropicScale(Transform):
"""Scales node positions by a randomly sampled factor ``s1, s2, s3``
within a given interval, *e.g.*, resulting in the following
transformation matrix
.. math::
\left[
\begin{array}{ccc}
s1 & 0 & 0 \\
0 & s2 & 0 \\
0 & 0 & s3 \\
\end{array}
\right]
for three-dimensional positions.
If the nodes have a `normal` attribute, we also reorient those
accordingly, while preserving their unit-norm.
Warning: any other absolute orientation-related attributes beside
`pos` and `normal` may be broken by this transform.
Credit: https://github.com/torch-points3d/torch-points3d
:param delta: float or List(float)
Scaling will be uniformly sampled in [-delta, delta]. If a
3-element list may be passed to scale X, Y and Z differently.
"""
_IN_TYPE = NAG
_OUT_TYPE = NAG
def __init__(self, delta=0.2):
assert isinstance(delta, (float, int)) or isinstance(delta, (tuple, list))
if isinstance(delta, (float, int)):
delta = [float(delta)] * 3
assert len(delta) == 3
self.delta = torch.tensor(delta).abs().view(1, -1)
def _process(self, nag):
# Generate the random scales
scale = 1 + (torch.rand(1) * 2 * self.delta - self.delta).to(nag.device)
for i_level in range(nag.num_levels):
nag[i_level].pos = nag[i_level].pos * scale
# If the nodes have a `normal` or 'mean_normal' attribute,
# we also adapt their orientations accordingly
for k in ['normal', 'mean_normal']:
if getattr(nag[i_level], k, None) is not None:
nag[i_level][k] = self._scale_normal(nag[i_level][k], scale)
# TODO: this is an ugly, hardcoded patch to deal with
# features assumedly created by
# _minimalistic_horizontal_edge_features........
if getattr(nag[i_level], 'edge_attr', None) is not None:
edge_attr = nag[i_level].edge_attr
assert edge_attr.shape[1] == 7, \
"Expected exactly 7 features in `edge_attr`, generated " \
"with `_minimalistic_horizontal_edge_features`"
edge_attr[:, :3] *= scale
edge_attr[:, 3:] *= scale.norm() # std_off and mean_dist are scaled by the scaling norm, slightly incorrect for std_off...
nag[i_level].edge_attr = edge_attr
return nag
@staticmethod
def _scale_normal(normal, scale):
return torch.nn.functional.normalize(normal * scale, dim=1)
class RandomAxisFlip(Transform):
"""Flip the node positions wrt one of the XYZ axes, with a specified
probability. This transform is not very modular because it is
intended to be composed with `RandomTiltAndRotate` for richer
geometric augmentations.
If the nodes have a `normal` or 'mean_normal' attribute, we also
flip those accordingly.
Warning: any other absolute orientation-related attributes beside
`pos`, `normal` 'mean_normal' may be broken by this transform.
:param p: float
Probability of flip
"""
_IN_TYPE = NAG
_OUT_TYPE = NAG
def __init__(self, axis=0, p=0.5):
assert isinstance(axis, int)
assert isinstance(p, float)
self.axis = axis
self.p = p
def _process(self, nag):
if torch.rand(1, device=nag.device) > self.p:
return nag
axis = self.axis
for i_level in range(nag.num_levels):
nag[i_level].pos[:, axis] *= -1
# If the nodes have a `normal` or 'mean_normal' attribute,
# we also adapt their orientations accordingly
for k in ['normal', 'mean_normal']:
if getattr(nag[i_level], k, None) is not None:
nag[i_level][k] = self._flip_normal(nag[i_level][k], axis)
# TODO: this is an ugly, hardcoded patch to deal with
# features assumedly created by
# _minimalistic_horizontal_edge_features........
if nag[i_level].edge_attr is not None:
edge_attr = nag[i_level].edge_attr
assert edge_attr.shape[1] == 7, \
"Expected exactly 7 features in `edge_attr`, generated " \
"with `_minimalistic_horizontal_edge_features`"
edge_attr[:, :3][:, axis] *= -1 # mean_off
nag[i_level].edge_attr = edge_attr
return nag
@staticmethod
def _flip_normal(normal, axis):
normal[:, axis] *= -1
normal[normal[:, 2] < 0] *= -1
return normal
|