File size: 11,826 Bytes
191d2cc | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 | """MEI (Motion Euclidean-Invariant) representation.
Converts between SMPL-H parameters and the 138-dimensional MEI vector,
which provides SE(2) invariance (ground-plane translation + yaw rotation).
MEI vector layout (138D):
[0:2] CoM planar velocity in local frame (v_x, v_z) 2D
[2:3] Heading angular velocity (d_theta / dt) 1D
[3:6] Root position relative to CoM pivot, local frame 3D
[6:12] Root rotation (heading-relative), 6D continuous 6D
[12:138] Body joint rotations, 21 joints x 6D 126D
---------------------------------------------------------------
Total 138D
Coordinate convention (local frame, after heading alignment):
X-axis: right
Y-axis: up
Z-axis: forward (heading direction)
"""
import numpy as np
from .geometry import (
axis_angle_to_matrix,
matrix_to_axis_angle,
matrix_to_rotation_6d,
rotation_6d_to_matrix,
yaw_rotation_matrix,
wrap_angle,
)
MEI_DIM = 138
NUM_JOINTS = 22 # root + 21 body joints
# SMPL 22-joint names
SMPL_JOINT_NAMES = [
"Pelvis", "L_Hip", "R_Hip", "Spine1",
"L_Knee", "R_Knee", "Spine2", "L_Ankle",
"R_Ankle", "Spine3", "L_Foot", "R_Foot",
"Neck", "L_Collar", "R_Collar", "Head",
"L_Shoulder", "R_Shoulder", "L_Elbow", "R_Elbow",
"L_Wrist", "R_Wrist",
]
# Approximate body-segment mass ratios for CoM computation.
# Adapted from de Leva (1996) for the SMPL 22-joint topology.
# Each weight approximates the fraction of total body mass best
# represented by the corresponding joint position.
# _MASS_WEIGHTS = np.array([
# 0.142, # 0 Pelvis
# 0.100, # 1 L_Hip
# 0.100, # 2 R_Hip
# 0.094, # 3 Spine1
# 0.047, # 4 L_Knee
# 0.047, # 5 R_Knee
# 0.094, # 6 Spine2
# 0.014, # 7 L_Ankle
# 0.014, # 8 R_Ankle
# 0.047, # 9 Spine3
# 0.014, # 10 L_Foot
# 0.014, # 11 R_Foot
# 0.028, # 12 Neck
# 0.019, # 13 L_Collar
# 0.019, # 14 R_Collar
# 0.066, # 15 Head
# 0.028, # 16 L_Shoulder
# 0.028, # 17 R_Shoulder
# 0.019, # 18 L_Elbow
# 0.019, # 19 R_Elbow
# 0.006, # 20 L_Wrist
# 0.006, # 21 R_Wrist
# ], dtype=np.float64)
_MASS_WEIGHTS = np.array([
1, # 0 Pelvis
0, # 1 L_Hip
0, # 2 R_Hip
0, # 3 Spine1
0, # 4 L_Knee
0, # 5 R_Knee
0, # 6 Spine2
0, # 7 L_Ankle
0, # 8 R_Ankle
0, # 9 Spine3
0, # 10 L_Foot
0, # 11 R_Foot
0, # 12 Neck
0, # 13 L_Collar
0, # 14 R_Collar
0, # 15 Head
0, # 16 L_Shoulder
0, # 17 R_Shoulder
0, # 18 L_Elbow
0, # 19 R_Elbow
0, # 20 L_Wrist
0, # 21 R_Wrist
], dtype=np.float64)
JOINT_MASS_WEIGHTS = _MASS_WEIGHTS / _MASS_WEIGHTS.sum()
# ------------------------------------------------------------------
# Helper functions
# ------------------------------------------------------------------
def compute_com(joints: np.ndarray) -> np.ndarray:
"""Compute center of mass from 22-joint positions.
Args:
joints: (T, 22, 3) world-frame joint positions.
Returns:
com: (T, 3) center of mass.
"""
w = JOINT_MASS_WEIGHTS.reshape(1, -1, 1) # (1, 22, 1)
return (joints * w).sum(axis=1)
def compute_com_ground(joints: np.ndarray) -> np.ndarray:
"""CoM projected onto the ground plane (Y = 0).
Args:
joints: (T, 22, 3) world-frame joint positions.
Returns:
com_ground: (T, 3) with Y component set to 0.
"""
com = compute_com(joints)
com[:, 1] = 0.0
return com
def compute_heading(joints: np.ndarray) -> np.ndarray:
"""Extract heading angle from joint positions (HumanML3D convention).
The heading (facing direction) is determined from the average of
hip and shoulder lateral (left-to-right) vectors, cross-producted
with the Y-up axis to obtain the forward direction in the XZ plane.
To avoid 180-degree flips, each frame's heading is compared with
the previous frame and the closer of {heading, heading + pi} is kept.
Args:
joints: (T, 22, 3) world-frame joint positions.
Returns:
heading: (T,) heading angles in radians.
"""
r_hip = SMPL_JOINT_NAMES.index("R_Hip")
l_hip = SMPL_JOINT_NAMES.index("L_Hip")
sdr_r = SMPL_JOINT_NAMES.index("R_Shoulder")
sdr_l = SMPL_JOINT_NAMES.index("L_Shoulder")
# Lateral vectors (left -> right)
across1 = joints[:, r_hip] - joints[:, l_hip] # (T, 3)
across2 = joints[:, sdr_r] - joints[:, sdr_l] # (T, 3)
across = across1 + across2 # (T, 3)
# Forward = Y_up x right_direction (result lies in XZ plane, Y=0)
# Do NOT normalize across beforehand: ||forward|| = ||across|| * sin(theta)
# naturally captures both degeneracies (across ≈ 0 and across ≈ vertical).
forward = np.cross(np.array([[0, 1, 0]]), across, axis=-1) # (T, 3)
xz_norm = np.linalg.norm(forward, axis=-1) # (T,)
forward = forward / np.maximum(xz_norm[:, np.newaxis], 1e-12)
raw_heading = np.arctan2(forward[:, 0], forward[:, 2])
# Resolve ambiguities frame by frame
heading = np.empty_like(raw_heading)
heading[0] = raw_heading[0]
for t in range(1, len(heading)):
# If across is nearly vertical, cross product is degenerate -> keep previous
if xz_norm[t] < 1e-3:
heading[t] = heading[t - 1]
continue
h = raw_heading[t]
h_flip = h + np.pi
# Pick whichever is closer to previous frame (resolve 180 flip)
diff_h = abs(wrap_angle(h - heading[t - 1]))
diff_flip = abs(wrap_angle(h_flip - heading[t - 1]))
heading[t] = h if diff_h <= diff_flip else wrap_angle(h_flip)
return heading
# ------------------------------------------------------------------
# Encoding: SMPL-H -> MEI
# ------------------------------------------------------------------
def smplh_to_mei(
root_orient: np.ndarray,
body_pose: np.ndarray,
joints: np.ndarray,
fps: float = 30.0,
) -> np.ndarray:
"""Convert SMPL-H parameters to MEI representation.
Args:
root_orient: (T, 3) root rotation in axis-angle.
body_pose: (T, 21, 3) or (T, 63) body joint rotations in axis-angle.
joints: (T, 22, 3) world-frame 3D joint positions (from SMPL-H FK).
joints[:, 0] is the pelvis (root) world position.
fps: frame rate (default 30).
Returns:
mei: (T, 138) MEI representation.
"""
T = root_orient.shape[0]
if body_pose.ndim == 2 and body_pose.shape[-1] == 63:
body_pose = body_pose.reshape(T, 21, 3)
# 1. Heading from joint positions (HumanML3D convention)
heading = compute_heading(joints) # (T,)
# 2. CoM ground projection
com_ground = compute_com_ground(joints) # (T, 3)
# 3. Heading angular velocity (backward difference)
heading_vel = np.zeros(T)
if T > 1:
heading_vel[0] = wrap_angle(heading[1] - heading[0]) * fps # forward difference for first frame
heading_vel[1:] = wrap_angle(heading[1:] - heading[:-1]) * fps
# 4. CoM planar velocity in local frame (backward diff)
R_heading = yaw_rotation_matrix(heading) # (T, 3, 3)
R_inv = np.transpose(R_heading, (0, 2, 1)) # (T, 3, 3)
com_vel_global = np.zeros_like(com_ground)
if T > 1:
com_vel_global[0] = (com_ground[1] - com_ground[0]) * fps # forward difference for first frame
com_vel_global[1:] = (com_ground[1:] - com_ground[:-1]) * fps
com_vel_local = np.einsum("tij,tj->ti", R_inv, com_vel_global) # (T, 3)
com_vel_planar = com_vel_local[:, [0, 2]] # (T, 2) [v_x, v_z]
# 5. Root position relative to CoM pivot, in local frame
root_offset_global = joints[:, 0] - com_ground # (T, 3)
root_offset_local = np.einsum("tij,tj->ti", R_inv, root_offset_global)
# 6. Root rotation relative to heading, encoded as 6D
root_rotmat = axis_angle_to_matrix(root_orient) # (T, 3, 3)
root_rotmat_local = np.einsum("tij,tjk->tik", R_inv, root_rotmat)
root_rot6d = matrix_to_rotation_6d(root_rotmat_local) # (T, 6)
# 7. Body joint rotations -> 6D (already in parent-local frame)
body_rotmat = axis_angle_to_matrix(body_pose) # (T, 21, 3, 3)
body_rot6d = matrix_to_rotation_6d(body_rotmat).reshape(T, -1) # (T, 126)
# Assemble
mei = np.concatenate([
com_vel_planar, # 0:2 (2)
heading_vel[:, None], # 2:3 (1)
root_offset_local, # 3:6 (3)
root_rot6d, # 6:12 (6)
body_rot6d, # 12:138 (126)
], axis=-1)
assert mei.shape == (T, MEI_DIM), f"Expected ({T}, {MEI_DIM}), got {mei.shape}"
return mei
# ------------------------------------------------------------------
# Decoding: MEI -> SMPL-H
# ------------------------------------------------------------------
def mei_to_smplh(
mei: np.ndarray,
fps: float = 30.0,
initial_heading: float = 0.0,
initial_com: np.ndarray = np.array([0.0, 0.0])
) -> dict:
"""Convert MEI representation back to SMPL-H parameters.
Because MEI is SE(2)-invariant, the absolute ground-plane position and
yaw are lost. They are recovered from *initial_heading* and
*initial_com* (both default to zero, placing the motion at the origin
facing +Z).
Args:
mei: (T, 138) MEI representation.
fps: frame rate (default 30).
initial_heading: heading angle at frame 0 (radians).
initial_com: (2,) XZ position of CoM at frame 0, default [0, 0].
Returns:
dict with:
root_orient: (T, 3) axis-angle root rotation.
body_pose: (T, 63) axis-angle body joint rotations.
transl: (T, 3) global root translation.
"""
T = mei.shape[0]
# --- Parse ---
com_vel_planar = mei[:, 0:2] # (T, 2)
heading_vel = mei[:, 2] # (T,)
root_offset_local = mei[:, 3:6] # (T, 3)
root_rot6d = mei[:, 6:12] # (T, 6)
body_rot6d = mei[:, 12:138] # (T, 126)
# 1. Integrate heading angular velocity
heading = np.zeros(T)
heading[0] = initial_heading
for t in range(1, T):
heading[t] = heading[t - 1] + heading_vel[t] / fps
heading = wrap_angle(heading)
# 2. Integrate CoM ground trajectory
R_heading = yaw_rotation_matrix(heading) # (T, 3, 3)
com_vel_local_3d = np.zeros((T, 3))
com_vel_local_3d[:, 0] = com_vel_planar[:, 0]
com_vel_local_3d[:, 2] = com_vel_planar[:, 1]
com_vel_global = np.einsum("tij,tj->ti", R_heading, com_vel_local_3d)
com_ground = np.zeros((T, 3))
com_ground[0, 0] = initial_com[0]
com_ground[0, 2] = initial_com[1]
for t in range(1, T):
com_ground[t] = com_ground[t - 1] + com_vel_global[t] / fps
# 3. Root position (global)
root_offset_global = np.einsum("tij,tj->ti", R_heading, root_offset_local)
transl = com_ground + root_offset_global
# 4. Root rotation (global)
root_rotmat_local = rotation_6d_to_matrix(root_rot6d) # (T, 3, 3)
root_rotmat = np.einsum("tij,tjk->tik", R_heading, root_rotmat_local)
root_orient = matrix_to_axis_angle(root_rotmat) # (T, 3)
# 5. Body joint rotations -> axis-angle
body_rotmat = rotation_6d_to_matrix(body_rot6d.reshape(T, 21, 6))
body_pose = matrix_to_axis_angle(body_rotmat).reshape(T, 63)
return {
"root_orient": root_orient,
"body_pose": body_pose,
"transl": transl,
}
|