Hugging Face's logo

Datasets:
mysigner
/
MySign

Tasks:
Other
Modalities:
3D
Text
Formats:
json
Languages:
Malaysian Sign Language
English
Size:
< 1K
Tags:
sign-language
bahasa-isyarat-malaysia
bim
malaysian-sign-language
motion-capture
3d
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
3
MySign / hand_angular_constraints.py
mysigner's picture
mysigner
Upload 2 files
38b191e verified
Raw History Blame
5.75 kB
#!/usr/bin/env python3
import torch
import os
import numpy as np
import trimesh
from common.utils.utils_hand import batch_euler2matzxy, coordtrans
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
euler_coordtrans_RIGHT = np.array([
[ 15, -6, 10,
15, 1, -5,
15, -15, 0,
0, 20, 2,
0, 10, -2,
0, 10, 10,
15, 25, 10,
10, 37, 8,
7, 35, 8,
0, 12, 5,
0, 25, -5,
0, 25, 15,
-30, -75, -35,
-30, -45, -30,
-30, -45, -30]]) / 180 * np.pi
euler_coordtrans_LEFT = np.array([
[ 15, 6, -10,
15, -1, 5,
15, 15, 0,
0, -20, -2,
0, -10, 2,
0, -10, -10,
15, -25, -10,
10, -37, -8,
7, -35, -8,
0, -12, -5,
0, -25, 5,
0, -25, -15,
-30, 75, 35,
-30, 45, 30,
-30, 45, 30]]) / 180 * np.pi
# range of motion
BOF_RIGHT = np.array([
[
0, -30, -40, # index0
0, 0, 0, # index1
0, 0, -5, # index2
0, -22.5, -40, # middle0
0, 0, 0, # middle1
0, 0, -5, # middle2
0, -25, -40, # pinky0
0, 0, 0, # pinky1
0, 0, -5, # pinky2
0, -22.5, -40, # ring0
0, 0, 0, # ring1
0, 0, -5, # ring2
-180, -60, -15, # thumb0
-180,-5, -180, # thumb1
-180, -5, -10], # thumb0
# MAX VALUES
[
0, 30, 90, # index0
0, 0, 110, # index1
0, 0, 90, # index2
0, 22.5, 90, # middle0
0, 0, 110, # middle1
0, 0, 90, # middle2
0, 25, 90, # pinky0
0, 0, 135, # pinky1
0, 0, 90, # pinky2
0, 22.5, 90, # ring0
0, 0, 120, # ring1
0, 0, 90, # ring2
180, 60, 90, # thumb0
180, 5, 80, # thumb1
180, 5, 80]])
BOF_LEFT = np.array([
[
0, -30, -90, # index0
0, 0,-110, # index1
0, 0, -90, # index2
0, -22.5, -90, # middle0
0, 0, -110, # middle1
0, 0, -90, # middle2
0, -25, -90, # pinky0
0, 0, -135, # pinky1
0, 0, -90, # pinky2
0, -22.5, -90, # ring0
0, 0, -120, # ring1
0, 0, -90, # ring2
-180, -60, -90, # thumb0
-180, -5, -80, # thumb1
-180, -5, -80], # thumb2
[
0, 30, 40, # index0
0, 0, 0, # index1
0, 0, 5, # index2
0, 22.5, 40, # middle0
0, 0, 0, # middle1
0, 0, 5, # middle2
0, 25, 40, # pinky0
0, 0, 0, # pinky1
0, 0, 5, # pinky2
0, 22.5, 40, # ring0
0, 0, 0, # ring1
0, 0, 5, # ring2
180, 60, 15, # thumb0
180, 5, 180, # thumb1
180, 5, 10], # thumb2
])
def do_clipping(input_angles, side, rotation_type):
device = input_angles.device
if side == "LEFT":
max = torch.tensor(BOF_LEFT[1,:], dtype=torch.float32).reshape(15,3).to(device)
min = torch.tensor(BOF_LEFT[0,:], dtype=torch.float32).reshape(15,3).to(device)
min_np = torch.deg2rad(min) # shape (N, 3)
max_np = torch.deg2rad(max)
elif side == "RIGHT":
max = torch.tensor(BOF_RIGHT[1,:], dtype=torch.float32).reshape(15,3).to(device)
min = torch.tensor(BOF_RIGHT[0,:], dtype=torch.float32).reshape(15,3).to(device)
min_np = torch.deg2rad(min) # shape (N, 3)
max_np = torch.deg2rad(max)
input_angles = input_angles.reshape(-1,15,3)
B, J, N = input_angles.shape
if rotation_type == "quat":
org_mats = quaternion_to_rotation_matrix(input_angles).float()
elif rotation_type == "aa":
org_mats = axis_angle_to_rotation_matrix(input_angles.reshape(-1, 3)).float()
if len(org_mats.shape)!=4:
org_mats = org_mats.unsqueeze(0)
rotmat_mano = org_mats.clone()
rotmat_mano = rotmat_mano.view(-1,15,3,3)
name = f"euler_coordtrans_{side.upper()}"
mat = globals()[name]
local2global = torch.from_numpy(mat).type(torch.float32).to(device)
local2global = batch_euler2matzxy(local2global.view(-1,3)).view(-1,15,3,3).expand(1,-1,-1,-1)
anatom_space_mats = coordtrans(rotmat_mano, local2global, 0)
mat_to_quat = rotation_matrix_to_quaternion(anatom_space_mats)
roll, pitch, yaw = euler_from_quaternion(mat_to_quat[:, :, 0], mat_to_quat[:, :, 1], mat_to_quat[:, :, 2], mat_to_quat[:, :, 3])
euler_from_quat = torch.stack([roll, pitch, yaw], dim=-1) # shape (1, 15, 3)
euler_from_quat = torch.clip(euler_from_quat, min_np, max_np) # root joint is always zero
qw, qx, qy, qz = quaternion_from_euler(euler_from_quat[:,:, 0], euler_from_quat[:,:, 1], euler_from_quat[:,:, 2])
rots_anat = torch.stack([qw, qx, qy, qz], dim=-1) # shape (1, 15, 4)
anat_mats = quaternion_to_rotation_matrix(rots_anat)
corrected_mano_space = coordtrans(anat_mats, local2global, 1)
corrected_aa = rotation_matrix_to_axis_angle(corrected_mano_space).reshape(B, J*3)
return corrected_aa