Spaces:
Running
Running
File size: 4,016 Bytes
b7eedf7 |
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 |
import torch
import torch.nn.functional as F
from lietorch import SE3, Sim3
MIN_DEPTH = 0.2
def extract_intrinsics(intrinsics):
return intrinsics[...,None,None,:].unbind(dim=-1)
def coords_grid(ht, wd, **kwargs):
y, x = torch.meshgrid(
torch.arange(ht).to(**kwargs).float(),
torch.arange(wd).to(**kwargs).float(), indexing='ij')
return torch.stack([x, y], dim=-1)
def iproj(disps, intrinsics, jacobian=False):
""" pinhole camera inverse projection """
ht, wd = disps.shape[2:]
fx, fy, cx, cy = extract_intrinsics(intrinsics)
y, x = torch.meshgrid(
torch.arange(ht).to(disps.device).float(),
torch.arange(wd).to(disps.device).float(), indexing='ij')
i = torch.ones_like(disps)
X = (x - cx) / fx
Y = (y - cy) / fy
pts = torch.stack([X, Y, i, disps], dim=-1)
if jacobian:
J = torch.zeros_like(pts)
J[...,-1] = 1.0
return pts, J
return pts, None
def proj(Xs, intrinsics, jacobian=False, return_depth=False):
""" pinhole camera projection """
fx, fy, cx, cy = extract_intrinsics(intrinsics)
X, Y, Z, D = Xs.unbind(dim=-1)
Z = torch.where(Z < 0.5*MIN_DEPTH, torch.ones_like(Z), Z)
d = 1.0 / Z
x = fx * (X * d) + cx
y = fy * (Y * d) + cy
if return_depth:
coords = torch.stack([x, y, D*d], dim=-1)
else:
coords = torch.stack([x, y], dim=-1)
if jacobian:
B, N, H, W = d.shape
o = torch.zeros_like(d)
proj_jac = torch.stack([
fx*d, o, -fx*X*d*d, o,
o, fy*d, -fy*Y*d*d, o,
# o, o, -D*d*d, d,
], dim=-1).view(B, N, H, W, 2, 4)
return coords, proj_jac
return coords, None
def actp(Gij, X0, jacobian=False):
""" action on point cloud """
X1 = Gij[:,:,None,None] * X0
if jacobian:
X, Y, Z, d = X1.unbind(dim=-1)
o = torch.zeros_like(d)
B, N, H, W = d.shape
if isinstance(Gij, SE3):
Ja = torch.stack([
d, o, o, o, Z, -Y,
o, d, o, -Z, o, X,
o, o, d, Y, -X, o,
o, o, o, o, o, o,
], dim=-1).view(B, N, H, W, 4, 6)
elif isinstance(Gij, Sim3):
Ja = torch.stack([
d, o, o, o, Z, -Y, X,
o, d, o, -Z, o, X, Y,
o, o, d, Y, -X, o, Z,
o, o, o, o, o, o, o
], dim=-1).view(B, N, H, W, 4, 7)
return X1, Ja
return X1, None
def projective_transform(poses, depths, intrinsics, ii, jj, jacobian=False, return_depth=False):
""" map points from ii->jj """
# inverse project (pinhole)
X0, Jz = iproj(depths[:,ii], intrinsics[:,ii], jacobian=jacobian)
# transform
Gij = poses[:,jj] * poses[:,ii].inv()
Gij.data[:,ii==jj] = torch.as_tensor([-0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], device="cuda")
X1, Ja = actp(Gij, X0, jacobian=jacobian)
# project (pinhole)
x1, Jp = proj(X1, intrinsics[:,jj], jacobian=jacobian, return_depth=return_depth)
# exclude points too close to camera
valid = ((X1[...,2] > MIN_DEPTH) & (X0[...,2] > MIN_DEPTH)).float()
valid = valid.unsqueeze(-1)
if jacobian:
# Ji transforms according to dual adjoint
Jj = torch.matmul(Jp, Ja)
Ji = -Gij[:,:,None,None,None].adjT(Jj)
Jz = Gij[:,:,None,None] * Jz
Jz = torch.matmul(Jp, Jz.unsqueeze(-1))
return x1, valid, (Ji, Jj, Jz)
return x1, valid
def induced_flow(poses, disps, intrinsics, ii, jj):
""" optical flow induced by camera motion """
ht, wd = disps.shape[2:]
y, x = torch.meshgrid(
torch.arange(ht).to(disps.device).float(),
torch.arange(wd).to(disps.device).float(), indexing='ij')
coords0 = torch.stack([x, y], dim=-1)
coords1, valid = projective_transform(poses, disps, intrinsics, ii, jj, False)
return coords1[...,:2] - coords0, valid
|