Upload 2 files
Browse files- body_angular_constraints.py +67 -0
- hand_angular_constraints.py +201 -0
body_angular_constraints.py
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from scipy.spatial.transform import Rotation as R
|
| 2 |
+
import torch, numpy as np
|
| 3 |
+
|
| 4 |
+
BOF_body = np.array([
|
| 5 |
+
[-120.0, -130.0, -80.0, # left shoulder
|
| 6 |
+
-120.0, 0.0, -80.0, # right shoulder
|
| 7 |
+
-180.0, -160.0, -180.0, # left elbow
|
| 8 |
+
-180.0, 0.0, -180.0, # right elbow
|
| 9 |
+
-120.0, -50.0, -90.0, # left wrist
|
| 10 |
+
-120.0, -50.0, -90.0, # right wrist
|
| 11 |
+
],
|
| 12 |
+
[90.0, 0.0, 80.0, # left shoulder
|
| 13 |
+
90.0, 130.0, 80.0, # right shoulder
|
| 14 |
+
180.0, 0.0, 180.0, # left elbow
|
| 15 |
+
180.0, 160.0, 180.0, # right elbow
|
| 16 |
+
90.0, 50.0, 90.0, # left wrist
|
| 17 |
+
90.0, 50.0, 90.0]]) / 180 * np.pi
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def _to_numpy_flat_last3(x: torch.Tensor):
|
| 21 |
+
dev, dt = x.device, x.dtype
|
| 22 |
+
x_np = x.detach().cpu().numpy().reshape(-1, 3)
|
| 23 |
+
return x_np, x.shape, dev, dt
|
| 24 |
+
|
| 25 |
+
def _from_numpy(x_np: np.ndarray, shape, dev, dt):
|
| 26 |
+
y = torch.from_numpy(x_np.reshape(*shape))
|
| 27 |
+
if dt in (torch.float32, torch.float64): y = y.to(dt)
|
| 28 |
+
return y.to(dev)
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def euler_XYZ_to_axis_angle_scipy(e: torch.Tensor, degrees: bool = False) -> torch.Tensor:
|
| 33 |
+
e_np, shape, dev, dt = _to_numpy_flat_last3(e)
|
| 34 |
+
aa_np = R.from_euler('XYZ', e_np, degrees=degrees).as_rotvec()
|
| 35 |
+
return _from_numpy(aa_np, shape, dev, dt)
|
| 36 |
+
|
| 37 |
+
def axis_angle_to_euler_XYZ_scipy(aa: torch.Tensor):
|
| 38 |
+
aa_np, shape, dev, dt = _to_numpy_flat_last3(aa)
|
| 39 |
+
e_np = R.from_rotvec(aa_np).as_euler('XYZ', degrees=False)
|
| 40 |
+
return _from_numpy(e_np, shape, dev, dt)
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def apply_angular_constraints(body_pose): # body_pose: (B, 63) axis-angle for 21 joints
|
| 46 |
+
device = body_pose.device
|
| 47 |
+
B = body_pose.shape[0]
|
| 48 |
+
body_pose = body_pose.view(B, 21, 3)
|
| 49 |
+
|
| 50 |
+
# your bounds (6×3) in radians, defined for joints [15..20] in intrinsic 'XYZ'
|
| 51 |
+
minC = torch.tensor(BOF_body[0], dtype=body_pose.dtype, device=device).view(6,3)
|
| 52 |
+
maxC = torch.tensor(BOF_body[1], dtype=body_pose.dtype, device=device).view(6,3)
|
| 53 |
+
|
| 54 |
+
aa_arms = body_pose[:, 15:, :]
|
| 55 |
+
|
| 56 |
+
e_arms = axis_angle_to_euler_XYZ_scipy(aa_arms) # (B,6,3) intrinsic 'XYZ'
|
| 57 |
+
|
| 58 |
+
e_clamp = torch.clamp(e_arms, minC, maxC)
|
| 59 |
+
|
| 60 |
+
aa_new = euler_XYZ_to_axis_angle_scipy(e_clamp)
|
| 61 |
+
|
| 62 |
+
body_pose[:, 15:, :] = aa_new
|
| 63 |
+
|
| 64 |
+
return body_pose.view(B, -1)
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
|
hand_angular_constraints.py
ADDED
|
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
import torch
|
| 3 |
+
import os
|
| 4 |
+
import numpy as np
|
| 5 |
+
import trimesh
|
| 6 |
+
from common.utils.utils_hand import batch_euler2matzxy, coordtrans
|
| 7 |
+
|
| 8 |
+
from kornia.geometry.conversions import axis_angle_to_rotation_matrix, quaternion_to_rotation_matrix, euler_from_quaternion, rotation_matrix_to_quaternion, quaternion_from_euler, rotation_matrix_to_axis_angle
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
euler_coordtrans_RIGHT = np.array([
|
| 12 |
+
[ 15, -6, 10,
|
| 13 |
+
15, 1, -5,
|
| 14 |
+
15, -15, 0,
|
| 15 |
+
0, 20, 2,
|
| 16 |
+
0, 10, -2,
|
| 17 |
+
0, 10, 10,
|
| 18 |
+
15, 25, 10,
|
| 19 |
+
10, 37, 8,
|
| 20 |
+
7, 35, 8,
|
| 21 |
+
0, 12, 5,
|
| 22 |
+
0, 25, -5,
|
| 23 |
+
0, 25, 15,
|
| 24 |
+
-30, -75, -35,
|
| 25 |
+
-30, -45, -30,
|
| 26 |
+
-30, -45, -30]]) / 180 * np.pi
|
| 27 |
+
euler_coordtrans_LEFT = np.array([
|
| 28 |
+
[ 15, 6, -10,
|
| 29 |
+
15, -1, 5,
|
| 30 |
+
15, 15, 0,
|
| 31 |
+
0, -20, -2,
|
| 32 |
+
0, -10, 2,
|
| 33 |
+
0, -10, -10,
|
| 34 |
+
15, -25, -10,
|
| 35 |
+
10, -37, -8,
|
| 36 |
+
7, -35, -8,
|
| 37 |
+
0, -12, -5,
|
| 38 |
+
0, -25, 5,
|
| 39 |
+
0, -25, -15,
|
| 40 |
+
-30, 75, 35,
|
| 41 |
+
-30, 45, 30,
|
| 42 |
+
-30, 45, 30]]) / 180 * np.pi
|
| 43 |
+
|
| 44 |
+
# range of motion
|
| 45 |
+
|
| 46 |
+
BOF_RIGHT = np.array([
|
| 47 |
+
[
|
| 48 |
+
0, -30, -40, # index0
|
| 49 |
+
0, 0, 0, # index1
|
| 50 |
+
0, 0, -5, # index2
|
| 51 |
+
|
| 52 |
+
0, -22.5, -40, # middle0
|
| 53 |
+
0, 0, 0, # middle1
|
| 54 |
+
0, 0, -5, # middle2
|
| 55 |
+
|
| 56 |
+
0, -25, -40, # pinky0
|
| 57 |
+
0, 0, 0, # pinky1
|
| 58 |
+
0, 0, -5, # pinky2
|
| 59 |
+
|
| 60 |
+
0, -22.5, -40, # ring0
|
| 61 |
+
0, 0, 0, # ring1
|
| 62 |
+
0, 0, -5, # ring2
|
| 63 |
+
|
| 64 |
+
-180, -60, -15, # thumb0
|
| 65 |
+
-180,-5, -180, # thumb1
|
| 66 |
+
-180, -5, -10], # thumb0
|
| 67 |
+
|
| 68 |
+
# MAX VALUES
|
| 69 |
+
[
|
| 70 |
+
|
| 71 |
+
0, 30, 90, # index0
|
| 72 |
+
0, 0, 110, # index1
|
| 73 |
+
0, 0, 90, # index2
|
| 74 |
+
|
| 75 |
+
0, 22.5, 90, # middle0
|
| 76 |
+
0, 0, 110, # middle1
|
| 77 |
+
0, 0, 90, # middle2
|
| 78 |
+
|
| 79 |
+
0, 25, 90, # pinky0
|
| 80 |
+
0, 0, 135, # pinky1
|
| 81 |
+
0, 0, 90, # pinky2
|
| 82 |
+
|
| 83 |
+
0, 22.5, 90, # ring0
|
| 84 |
+
0, 0, 120, # ring1
|
| 85 |
+
0, 0, 90, # ring2
|
| 86 |
+
|
| 87 |
+
180, 60, 90, # thumb0
|
| 88 |
+
180, 5, 80, # thumb1
|
| 89 |
+
180, 5, 80]])
|
| 90 |
+
|
| 91 |
+
BOF_LEFT = np.array([
|
| 92 |
+
[
|
| 93 |
+
|
| 94 |
+
0, -30, -90, # index0
|
| 95 |
+
0, 0,-110, # index1
|
| 96 |
+
0, 0, -90, # index2
|
| 97 |
+
|
| 98 |
+
0, -22.5, -90, # middle0
|
| 99 |
+
0, 0, -110, # middle1
|
| 100 |
+
0, 0, -90, # middle2
|
| 101 |
+
|
| 102 |
+
0, -25, -90, # pinky0
|
| 103 |
+
0, 0, -135, # pinky1
|
| 104 |
+
0, 0, -90, # pinky2
|
| 105 |
+
|
| 106 |
+
0, -22.5, -90, # ring0
|
| 107 |
+
0, 0, -120, # ring1
|
| 108 |
+
0, 0, -90, # ring2
|
| 109 |
+
|
| 110 |
+
-180, -60, -90, # thumb0
|
| 111 |
+
-180, -5, -80, # thumb1
|
| 112 |
+
-180, -5, -80], # thumb2
|
| 113 |
+
|
| 114 |
+
[
|
| 115 |
+
|
| 116 |
+
0, 30, 40, # index0
|
| 117 |
+
0, 0, 0, # index1
|
| 118 |
+
0, 0, 5, # index2
|
| 119 |
+
|
| 120 |
+
0, 22.5, 40, # middle0
|
| 121 |
+
0, 0, 0, # middle1
|
| 122 |
+
0, 0, 5, # middle2
|
| 123 |
+
|
| 124 |
+
0, 25, 40, # pinky0
|
| 125 |
+
0, 0, 0, # pinky1
|
| 126 |
+
0, 0, 5, # pinky2
|
| 127 |
+
|
| 128 |
+
0, 22.5, 40, # ring0
|
| 129 |
+
0, 0, 0, # ring1
|
| 130 |
+
0, 0, 5, # ring2
|
| 131 |
+
|
| 132 |
+
180, 60, 15, # thumb0
|
| 133 |
+
180, 5, 180, # thumb1
|
| 134 |
+
180, 5, 10], # thumb2
|
| 135 |
+
])
|
| 136 |
+
|
| 137 |
+
def do_clipping(input_angles, side, rotation_type):
|
| 138 |
+
|
| 139 |
+
device = input_angles.device
|
| 140 |
+
|
| 141 |
+
if side == "LEFT":
|
| 142 |
+
max = torch.tensor(BOF_LEFT[1,:], dtype=torch.float32).reshape(15,3).to(device)
|
| 143 |
+
min = torch.tensor(BOF_LEFT[0,:], dtype=torch.float32).reshape(15,3).to(device)
|
| 144 |
+
min_np = torch.deg2rad(min) # shape (N, 3)
|
| 145 |
+
max_np = torch.deg2rad(max)
|
| 146 |
+
|
| 147 |
+
elif side == "RIGHT":
|
| 148 |
+
max = torch.tensor(BOF_RIGHT[1,:], dtype=torch.float32).reshape(15,3).to(device)
|
| 149 |
+
min = torch.tensor(BOF_RIGHT[0,:], dtype=torch.float32).reshape(15,3).to(device)
|
| 150 |
+
min_np = torch.deg2rad(min) # shape (N, 3)
|
| 151 |
+
max_np = torch.deg2rad(max)
|
| 152 |
+
|
| 153 |
+
input_angles = input_angles.reshape(-1,15,3)
|
| 154 |
+
|
| 155 |
+
B, J, N = input_angles.shape
|
| 156 |
+
|
| 157 |
+
if rotation_type == "quat":
|
| 158 |
+
|
| 159 |
+
org_mats = quaternion_to_rotation_matrix(input_angles).float()
|
| 160 |
+
|
| 161 |
+
elif rotation_type == "aa":
|
| 162 |
+
|
| 163 |
+
org_mats = axis_angle_to_rotation_matrix(input_angles.reshape(-1, 3)).float()
|
| 164 |
+
|
| 165 |
+
if len(org_mats.shape)!=4:
|
| 166 |
+
|
| 167 |
+
org_mats = org_mats.unsqueeze(0)
|
| 168 |
+
|
| 169 |
+
rotmat_mano = org_mats.clone()
|
| 170 |
+
|
| 171 |
+
rotmat_mano = rotmat_mano.view(-1,15,3,3)
|
| 172 |
+
|
| 173 |
+
name = f"euler_coordtrans_{side.upper()}"
|
| 174 |
+
mat = globals()[name]
|
| 175 |
+
|
| 176 |
+
local2global = torch.from_numpy(mat).type(torch.float32).to(device)
|
| 177 |
+
|
| 178 |
+
local2global = batch_euler2matzxy(local2global.view(-1,3)).view(-1,15,3,3).expand(1,-1,-1,-1)
|
| 179 |
+
|
| 180 |
+
anatom_space_mats = coordtrans(rotmat_mano, local2global, 0)
|
| 181 |
+
|
| 182 |
+
mat_to_quat = rotation_matrix_to_quaternion(anatom_space_mats)
|
| 183 |
+
|
| 184 |
+
roll, pitch, yaw = euler_from_quaternion(mat_to_quat[:, :, 0], mat_to_quat[:, :, 1], mat_to_quat[:, :, 2], mat_to_quat[:, :, 3])
|
| 185 |
+
|
| 186 |
+
euler_from_quat = torch.stack([roll, pitch, yaw], dim=-1) # shape (1, 15, 3)
|
| 187 |
+
|
| 188 |
+
euler_from_quat = torch.clip(euler_from_quat, min_np, max_np) # root joint is always zero
|
| 189 |
+
|
| 190 |
+
qw, qx, qy, qz = quaternion_from_euler(euler_from_quat[:,:, 0], euler_from_quat[:,:, 1], euler_from_quat[:,:, 2])
|
| 191 |
+
|
| 192 |
+
rots_anat = torch.stack([qw, qx, qy, qz], dim=-1) # shape (1, 15, 4)
|
| 193 |
+
|
| 194 |
+
anat_mats = quaternion_to_rotation_matrix(rots_anat)
|
| 195 |
+
|
| 196 |
+
corrected_mano_space = coordtrans(anat_mats, local2global, 1)
|
| 197 |
+
|
| 198 |
+
corrected_aa = rotation_matrix_to_axis_angle(corrected_mano_space).reshape(B, J*3)
|
| 199 |
+
|
| 200 |
+
return corrected_aa
|
| 201 |
+
|