File size: 5,311 Bytes
801501a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from .. import constants
import functools
from scipy.spatial.transform.rotation import Rotation
from ..custom_types import *


def quat_to_rot(q):
    shape = q.shape
    q = q.view(-1, 4)
    q_sq = 2 * q[:, :, None] * q[:, None, :]
    m00 = 1 - q_sq[:, 1, 1] - q_sq[:, 2, 2]
    m01 = q_sq[:, 0, 1] - q_sq[:, 2, 3]
    m02 = q_sq[:, 0, 2] + q_sq[:, 1, 3]

    m10 = q_sq[:, 0, 1] + q_sq[:, 2, 3]
    m11 = 1 - q_sq[:, 0, 0] - q_sq[:, 2, 2]
    m12 = q_sq[:, 1, 2] - q_sq[:, 0, 3]

    m20 = q_sq[:, 0, 2] - q_sq[:, 1, 3]
    m21 = q_sq[:, 1, 2] + q_sq[:, 0, 3]
    m22 = 1 - q_sq[:, 0, 0] - q_sq[:, 1, 1]
    r = torch.stack((m00, m01, m02, m10, m11, m12, m20, m21, m22), dim=1)
    r = r.view(*shape[:-1], 3, 3)
    return r


def rot_to_quat(r):
    shape = r.shape
    r = r.view(-1, 3, 3)
    qw = .5 * (1 + r[:, 0, 0] + r[:, 1, 1] + r[:, 2, 2]).sqrt()
    qx = (r[:, 2, 1] - r[:, 1, 2]) / (4 * qw)
    qy = (r[:, 0, 2] - r[:, 2, 0]) / (4 * qw)
    qz = (r[:, 1, 0] - r[:, 0, 1]) / (4 * qw)
    q = torch.stack((qx, qy, qz, qw), -1)
    q = q.view(*shape[:-2], 4)
    return q


@functools.lru_cache(10)
def get_rotation_matrix(theta: float, axis: float, degree: bool = False) -> ARRAY:
    if degree:
        theta = theta * np.pi / 180
    rotate_mat = np.eye(3)
    rotate_mat[axis, axis] = 1
    cos_theta, sin_theta = np.cos(theta), np.sin(theta)
    rotate_mat[(axis + 1) % 3, (axis + 1) % 3] = cos_theta
    rotate_mat[(axis + 2) % 3, (axis + 2) % 3] = cos_theta
    rotate_mat[(axis + 1) % 3, (axis + 2) % 3] = sin_theta
    rotate_mat[(axis + 2) % 3, (axis + 1) % 3] = -sin_theta
    return rotate_mat


def get_random_rotation(batch_size: int) -> T:
    r = Rotation.random(batch_size).as_matrix().astype(np.float32)
    Rotation.random()
    return torch.from_numpy(r)


def rand_bounded_rotation_matrix(cache_size: int, theta_range: float = .1):

    def create_cache():
        # from http://www.realtimerendering.com/resources/GraphicsGems/gemsiii/rand_rotation.c
        with torch.no_grad():
            theta, phi, z = torch.rand(cache_size, 3).split((1, 1, 1), dim=1)
            theta = (2 * theta - 1) * theta_range + 1
            theta = np.pi * theta   # Rotation about the pole (Z).
            phi = phi * 2 * np.pi  # For direction of pole deflection.
            z = 2 * z * theta_range # For magnitude of pole deflection.
            r = z.sqrt()
            v = torch.cat((torch.sin(phi) * r, torch.cos(phi) * r, torch.sqrt(2.0 - z)), dim=1)
            st = torch.sin(theta).squeeze(1)
            ct = torch.cos(theta).squeeze(1)
            rot_ = torch.zeros(cache_size, 3, 3)
            rot_[:, 0, 0] = ct
            rot_[:, 1, 1] = ct
            rot_[:, 0, 1] = st
            rot_[:, 1, 0] = -st
            rot_[:, 2, 2] = 1
            rot = (torch.einsum('ba,bd->bad', v, v) - torch.eye(3)[None, :, :]).bmm(rot_)
            det = rot.det()
        assert (det.gt(0.99) * det.lt(1.0001)).all().item()
        return rot

    def get_batch_rot(batch_size):
        nonlocal cache
        select = torch.randint(cache_size, size=(batch_size,))
        return cache[select]

    cache = create_cache()

    return get_batch_rot


def transform_rotation(points: T, one_axis=False, max_angle=-1):
    r = get_random_rotation(one_axis, max_angle)
    transformed = torch.einsum('nd,rd->nr', points, r)
    return transformed


def tb_to_rot(abc: T) -> T:
    c, s = torch.cos(abc), torch.sin(abc)
    aa = c[:, 0] * c[:, 1]
    ab = c[:, 0] * s[:, 1] * s[:, 2] - c[:, 2] * s[:, 0]
    ac = s[:, 0] * s[:, 2] + c[:, 0] * c[:, 2] * s[:, 1]

    ba = c[:, 1] * s[:, 0]
    bb = c[:, 0] * c[:, 2] + s.prod(-1)
    bc = c[:, 2] * s[:, 0] * s[:, 1] - c[:, 0] * s[:, 2]

    ca = -s[:, 1]
    cb = c[:, 1] * s[:, 2]
    cc = c[:, 1] * c[:, 2]
    return torch.stack((aa, ab, ac, ba, bb, bc, ca, cb, cc), 1).view(-1, 3, 3)


def rot_to_tb(rot: T) -> T:
    sy = torch.sqrt(rot[:, 0, 0] * rot[:, 0, 0] + rot[:, 1, 0] * rot[:, 1, 0])
    out = torch.zeros(rot.shape[0], 3, device = rot.device)
    mask = sy.gt(1e-6)
    z = torch.atan2(rot[mask, 2, 1], rot[mask, 2, 2])
    y = torch.atan2(-rot[mask, 2, 0], sy[mask])
    x = torch.atan2(rot[mask, 1, 0], rot[mask, 0, 0])
    out[mask] = torch.stack((x, y, z), dim=1)
    if not mask.all():
        mask = ~mask
        z = torch.atan2(-rot[mask, 1, 2], rot[mask, 1, 1])
        y = torch.atan2(-rot[mask, 2, 0], sy[mask])
        x = torch.zeros(x.shape)
        out[mask] = torch.stack((x, y, z), dim=1)
    return out


def apply_gmm_affine(gmms: TS, affine: T):
    mu, p, phi, eigen = gmms
    if affine.dim() == 2:
        affine = affine.unsqueeze(0).expand(mu.shape[0], *affine.shape)
    mu_r = torch.einsum('bad, bpnd->bpna', affine, mu)
    p_r = torch.einsum('bad, bpncd->bpnca', affine, p)
    return mu_r, p_r, phi, eigen


def get_reflection(reflect_axes: Tuple[bool, ...]) -> T:
    reflect = torch.eye(constants.DIM)
    for i in range(constants.DIM):
        if reflect_axes[i]:
            reflect[i, i] = -1
    return reflect


def get_tait_bryan_from_p(p: T) -> T:
    # p = p.squeeze(1)
    shape = p.shape
    rot = p.reshape(-1, 3, 3).permute(0, 2, 1)
    angles = rot_to_tb(rot)
    angles = angles / np.pi
    angles[:, 1] = angles[:, 1] * 2
    angles = angles.view(*shape[:2], 3)
    return angles