File size: 10,144 Bytes
62c7319 8b973ee 62c7319 8b973ee 62c7319 8b973ee 62c7319 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 |
import numpy as np
import cv2
import torch
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode
import torch.nn.functional as F
from PIL import Image
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Code taken from https://github.com/PruneTruong/DenseMatching/blob/40c29a6b5c35e86b9509e65ab0cd12553d998e5f/validation/utils_pose_estimation.py
# --- GEOMETRY ---
def estimate_pose(kpts0, kpts1, K0, K1, norm_thresh, conf=0.99999):
if len(kpts0) < 5:
return None
K0inv = np.linalg.inv(K0[:2, :2])
K1inv = np.linalg.inv(K1[:2, :2])
kpts0 = (K0inv @ (kpts0 - K0[None, :2, 2]).T).T
kpts1 = (K1inv @ (kpts1 - K1[None, :2, 2]).T).T
E, mask = cv2.findEssentialMat(
kpts0, kpts1, np.eye(3), threshold=norm_thresh, prob=conf, method=cv2.RANSAC
)
ret = None
if E is not None:
best_num_inliers = 0
for _E in np.split(E, len(E) / 3):
n, R, t, _ = cv2.recoverPose(_E, kpts0, kpts1, np.eye(3), 1e9, mask=mask)
if n > best_num_inliers:
best_num_inliers = n
ret = (R, t, mask.ravel() > 0)
return ret
def rotate_intrinsic(K, n):
base_rot = np.array([[0, 1, 0], [-1, 0, 0], [0, 0, 1]])
rot = np.linalg.matrix_power(base_rot, n)
return rot @ K
def rotate_pose_inplane(i_T_w, rot):
rotation_matrices = [
np.array(
[
[np.cos(r), -np.sin(r), 0.0, 0.0],
[np.sin(r), np.cos(r), 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 0.0, 1.0],
],
dtype=np.float32,
)
for r in [np.deg2rad(d) for d in (0, 270, 180, 90)]
]
return np.dot(rotation_matrices[rot], i_T_w)
def scale_intrinsics(K, scales):
scales = np.diag([1.0 / scales[0], 1.0 / scales[1], 1.0])
return np.dot(scales, K)
def to_homogeneous(points):
return np.concatenate([points, np.ones_like(points[:, :1])], axis=-1)
def angle_error_mat(R1, R2):
cos = (np.trace(np.dot(R1.T, R2)) - 1) / 2
cos = np.clip(cos, -1.0, 1.0) # numercial errors can make it out of bounds
return np.rad2deg(np.abs(np.arccos(cos)))
def angle_error_vec(v1, v2):
n = np.linalg.norm(v1) * np.linalg.norm(v2)
return np.rad2deg(np.arccos(np.clip(np.dot(v1, v2) / n, -1.0, 1.0)))
def compute_pose_error(T_0to1, R, t):
R_gt = T_0to1[:3, :3]
t_gt = T_0to1[:3, 3]
error_t = angle_error_vec(t.squeeze(), t_gt)
error_t = np.minimum(error_t, 180 - error_t) # ambiguity of E estimation
error_R = angle_error_mat(R, R_gt)
return error_t, error_R
def pose_auc(errors, thresholds):
sort_idx = np.argsort(errors)
errors = np.array(errors.copy())[sort_idx]
recall = (np.arange(len(errors)) + 1) / len(errors)
errors = np.r_[0.0, errors]
recall = np.r_[0.0, recall]
aucs = []
for t in thresholds:
last_index = np.searchsorted(errors, t)
r = np.r_[recall[:last_index], recall[last_index - 1]]
e = np.r_[errors[:last_index], t]
aucs.append(np.trapz(r, x=e) / t)
return aucs
# From Patch2Pix https://github.com/GrumpyZhou/patch2pix
def get_depth_tuple_transform_ops(resize=None, normalize=True, unscale=False):
ops = []
if resize:
ops.append(TupleResize(resize, mode=InterpolationMode.BILINEAR))
return TupleCompose(ops)
def get_tuple_transform_ops(resize=None, normalize=True, unscale=False):
ops = []
if resize:
ops.append(TupleResize(resize))
if normalize:
ops.append(TupleToTensorScaled())
ops.append(
TupleNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
) # Imagenet mean/std
else:
if unscale:
ops.append(TupleToTensorUnscaled())
else:
ops.append(TupleToTensorScaled())
return TupleCompose(ops)
class ToTensorScaled(object):
"""Convert a RGB PIL Image to a CHW ordered Tensor, scale the range to [0, 1]"""
def __call__(self, im):
if not isinstance(im, torch.Tensor):
im = np.array(im, dtype=np.float32).transpose((2, 0, 1))
im /= 255.0
return torch.from_numpy(im)
else:
return im
def __repr__(self):
return "ToTensorScaled(./255)"
class TupleToTensorScaled(object):
def __init__(self):
self.to_tensor = ToTensorScaled()
def __call__(self, im_tuple):
return [self.to_tensor(im) for im in im_tuple]
def __repr__(self):
return "TupleToTensorScaled(./255)"
class ToTensorUnscaled(object):
"""Convert a RGB PIL Image to a CHW ordered Tensor"""
def __call__(self, im):
return torch.from_numpy(np.array(im, dtype=np.float32).transpose((2, 0, 1)))
def __repr__(self):
return "ToTensorUnscaled()"
class TupleToTensorUnscaled(object):
"""Convert a RGB PIL Image to a CHW ordered Tensor"""
def __init__(self):
self.to_tensor = ToTensorUnscaled()
def __call__(self, im_tuple):
return [self.to_tensor(im) for im in im_tuple]
def __repr__(self):
return "TupleToTensorUnscaled()"
class TupleResize(object):
def __init__(self, size, mode=InterpolationMode.BICUBIC):
self.size = size
self.resize = transforms.Resize(size, mode)
def __call__(self, im_tuple):
return [self.resize(im) for im in im_tuple]
def __repr__(self):
return "TupleResize(size={})".format(self.size)
class TupleNormalize(object):
def __init__(self, mean, std):
self.mean = mean
self.std = std
self.normalize = transforms.Normalize(mean=mean, std=std)
def __call__(self, im_tuple):
return [self.normalize(im) for im in im_tuple]
def __repr__(self):
return "TupleNormalize(mean={}, std={})".format(self.mean, self.std)
class TupleCompose(object):
def __init__(self, transforms):
self.transforms = transforms
def __call__(self, im_tuple):
for t in self.transforms:
im_tuple = t(im_tuple)
return im_tuple
def __repr__(self):
format_string = self.__class__.__name__ + "("
for t in self.transforms:
format_string += "\n"
format_string += " {0}".format(t)
format_string += "\n)"
return format_string
@torch.no_grad()
def warp_kpts(kpts0, depth0, depth1, T_0to1, K0, K1):
"""Warp kpts0 from I0 to I1 with depth, K and Rt
Also check covisibility and depth consistency.
Depth is consistent if relative error < 0.2 (hard-coded).
# https://github.com/zju3dv/LoFTR/blob/94e98b695be18acb43d5d3250f52226a8e36f839/src/loftr/utils/geometry.py adapted from here
Args:
kpts0 (torch.Tensor): [N, L, 2] - <x, y>, should be normalized in (-1,1)
depth0 (torch.Tensor): [N, H, W],
depth1 (torch.Tensor): [N, H, W],
T_0to1 (torch.Tensor): [N, 3, 4],
K0 (torch.Tensor): [N, 3, 3],
K1 (torch.Tensor): [N, 3, 3],
Returns:
calculable_mask (torch.Tensor): [N, L]
warped_keypoints0 (torch.Tensor): [N, L, 2] <x0_hat, y1_hat>
"""
(
n,
h,
w,
) = depth0.shape
kpts0_depth = F.grid_sample(depth0[:, None], kpts0[:, :, None], mode="bilinear")[
:, 0, :, 0
]
kpts0 = torch.stack(
(w * (kpts0[..., 0] + 1) / 2, h * (kpts0[..., 1] + 1) / 2), dim=-1
) # [-1+1/h, 1-1/h] -> [0.5, h-0.5]
# Sample depth, get calculable_mask on depth != 0
nonzero_mask = kpts0_depth != 0
# Unproject
kpts0_h = (
torch.cat([kpts0, torch.ones_like(kpts0[:, :, [0]])], dim=-1)
* kpts0_depth[..., None]
) # (N, L, 3)
kpts0_n = K0.inverse() @ kpts0_h.transpose(2, 1) # (N, 3, L)
kpts0_cam = kpts0_n
# Rigid Transform
w_kpts0_cam = T_0to1[:, :3, :3] @ kpts0_cam + T_0to1[:, :3, [3]] # (N, 3, L)
w_kpts0_depth_computed = w_kpts0_cam[:, 2, :]
# Project
w_kpts0_h = (K1 @ w_kpts0_cam).transpose(2, 1) # (N, L, 3)
w_kpts0 = w_kpts0_h[:, :, :2] / (
w_kpts0_h[:, :, [2]] + 1e-4
) # (N, L, 2), +1e-4 to avoid zero depth
# Covisible Check
h, w = depth1.shape[1:3]
covisible_mask = (
(w_kpts0[:, :, 0] > 0)
* (w_kpts0[:, :, 0] < w - 1)
* (w_kpts0[:, :, 1] > 0)
* (w_kpts0[:, :, 1] < h - 1)
)
w_kpts0 = torch.stack(
(2 * w_kpts0[..., 0] / w - 1, 2 * w_kpts0[..., 1] / h - 1), dim=-1
) # from [0.5,h-0.5] -> [-1+1/h, 1-1/h]
# w_kpts0[~covisible_mask, :] = -5 # xd
w_kpts0_depth = F.grid_sample(
depth1[:, None], w_kpts0[:, :, None], mode="bilinear"
)[:, 0, :, 0]
consistent_mask = (
(w_kpts0_depth - w_kpts0_depth_computed) / w_kpts0_depth
).abs() < 0.05
valid_mask = nonzero_mask * covisible_mask * consistent_mask
return valid_mask, w_kpts0
imagenet_mean = torch.tensor([0.485, 0.456, 0.406]).to(device)
imagenet_std = torch.tensor([0.229, 0.224, 0.225]).to(device)
def numpy_to_pil(x: np.ndarray):
"""
Args:
x: Assumed to be of shape (h,w,c)
"""
if isinstance(x, torch.Tensor):
x = x.detach().cpu().numpy()
if x.max() <= 1.01:
x *= 255
x = x.astype(np.uint8)
return Image.fromarray(x)
def tensor_to_pil(x, unnormalize=False):
if unnormalize:
x = x * imagenet_std[:, None, None] + imagenet_mean[:, None, None]
x = x.detach().permute(1, 2, 0).cpu().numpy()
x = np.clip(x, 0.0, 1.0)
return numpy_to_pil(x)
def to_cuda(batch):
for key, value in batch.items():
if isinstance(value, torch.Tensor):
batch[key] = value.to(device)
return batch
def to_cpu(batch):
for key, value in batch.items():
if isinstance(value, torch.Tensor):
batch[key] = value.cpu()
return batch
def get_pose(calib):
w, h = np.array(calib["imsize"])[0]
return np.array(calib["K"]), np.array(calib["R"]), np.array(calib["T"]).T, h, w
def compute_relative_pose(R1, t1, R2, t2):
rots = R2 @ (R1.T)
trans = -rots @ t1 + t2
return rots, trans
|