Spaces:
Configuration error
Configuration error
File size: 14,062 Bytes
1ba539f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 |
import numpy as np
from lib.utils import base_utils
import cv2
from lib.config import cfg
import trimesh
def get_rays(H, W, K, R, T):
# calculate the camera origin
rays_o = -np.dot(R.T, T).ravel()
# calculate the world coodinates of pixels
i, j = np.meshgrid(np.arange(W, dtype=np.float32),
np.arange(H, dtype=np.float32),
indexing='xy')
xy1 = np.stack([i, j, np.ones_like(i)], axis=2)
pixel_camera = np.dot(xy1, np.linalg.inv(K).T)
pixel_world = np.dot(pixel_camera - T.ravel(), R)
# calculate the ray direction
rays_d = pixel_world - rays_o[None, None]
rays_o = np.broadcast_to(rays_o, rays_d.shape)
return rays_o, rays_d
def get_bound_corners(bounds):
min_x, min_y, min_z = bounds[0]
max_x, max_y, max_z = bounds[1]
corners_3d = np.array([
[min_x, min_y, min_z],
[min_x, min_y, max_z],
[min_x, max_y, min_z],
[min_x, max_y, max_z],
[max_x, min_y, min_z],
[max_x, min_y, max_z],
[max_x, max_y, min_z],
[max_x, max_y, max_z],
])
return corners_3d
def get_bound_2d_mask(bounds, K, pose, H, W):
corners_3d = get_bound_corners(bounds)
corners_2d = base_utils.project(corners_3d, K, pose)
corners_2d = np.round(corners_2d).astype(int)
mask = np.zeros((H, W), dtype=np.uint8)
cv2.fillPoly(mask, [corners_2d[[0, 1, 3, 2, 0]]], 1)
cv2.fillPoly(mask, [corners_2d[[4, 5, 7, 6, 5]]], 1)
cv2.fillPoly(mask, [corners_2d[[0, 1, 5, 4, 0]]], 1)
cv2.fillPoly(mask, [corners_2d[[2, 3, 7, 6, 2]]], 1)
cv2.fillPoly(mask, [corners_2d[[0, 2, 6, 4, 0]]], 1)
cv2.fillPoly(mask, [corners_2d[[1, 3, 7, 5, 1]]], 1)
return mask
def get_near_far(bounds, ray_o, ray_d):
"""calculate intersections with 3d bounding box"""
norm_d = np.linalg.norm(ray_d, axis=-1, keepdims=True)
viewdir = ray_d / norm_d
viewdir[(viewdir < 1e-5) & (viewdir > -1e-10)] = 1e-5
viewdir[(viewdir > -1e-5) & (viewdir < 1e-10)] = -1e-5
tmin = (bounds[:1] - ray_o[:1]) / viewdir
tmax = (bounds[1:2] - ray_o[:1]) / viewdir
t1 = np.minimum(tmin, tmax)
t2 = np.maximum(tmin, tmax)
near = np.max(t1, axis=-1)
far = np.min(t2, axis=-1)
mask_at_box = near < far
near = near[mask_at_box] / norm_d[mask_at_box, 0]
far = far[mask_at_box] / norm_d[mask_at_box, 0]
return near, far, mask_at_box
def sample_ray(img, msk, K, R, T, bounds, nrays, split):
H, W = img.shape[:2]
ray_o, ray_d = get_rays(H, W, K, R, T)
pose = np.concatenate([R, T], axis=1)
bound_mask = get_bound_2d_mask(bounds, K, pose, H, W)
msk = msk * bound_mask
if split == 'train':
nsampled_rays = 0
face_sample_ratio = cfg.face_sample_ratio
body_sample_ratio = cfg.body_sample_ratio
ray_o_list = []
ray_d_list = []
rgb_list = []
near_list = []
far_list = []
coord_list = []
mask_at_box_list = []
while nsampled_rays < nrays:
n_body = int((nrays - nsampled_rays) * body_sample_ratio)
n_face = int((nrays - nsampled_rays) * face_sample_ratio)
n_rand = (nrays - nsampled_rays) - n_body - n_face
# sample rays on body
coord_body = np.argwhere(msk != 0)
coord_body = coord_body[np.random.randint(0, len(coord_body),
n_body)]
# sample rays on face
coord_face = np.argwhere(msk == 13)
if len(coord_face) > 0:
coord_face = coord_face[np.random.randint(
0, len(coord_face), n_face)]
# sample rays in the bound mask
coord = np.argwhere(bound_mask == 1)
coord = coord[np.random.randint(0, len(coord), n_rand)]
if len(coord_face) > 0:
coord = np.concatenate([coord_body, coord_face, coord], axis=0)
else:
coord = np.concatenate([coord_body, coord], axis=0)
ray_o_ = ray_o[coord[:, 0], coord[:, 1]]
ray_d_ = ray_d[coord[:, 0], coord[:, 1]]
rgb_ = img[coord[:, 0], coord[:, 1]]
near_, far_, mask_at_box = get_near_far(bounds, ray_o_, ray_d_)
ray_o_list.append(ray_o_[mask_at_box])
ray_d_list.append(ray_d_[mask_at_box])
rgb_list.append(rgb_[mask_at_box])
near_list.append(near_)
far_list.append(far_)
coord_list.append(coord[mask_at_box])
mask_at_box_list.append(mask_at_box[mask_at_box])
nsampled_rays += len(near_)
ray_o = np.concatenate(ray_o_list).astype(np.float32)
ray_d = np.concatenate(ray_d_list).astype(np.float32)
rgb = np.concatenate(rgb_list).astype(np.float32)
near = np.concatenate(near_list).astype(np.float32)
far = np.concatenate(far_list).astype(np.float32)
coord = np.concatenate(coord_list)
mask_at_box = np.concatenate(mask_at_box_list)
else:
rgb = img.reshape(-1, 3).astype(np.float32)
ray_o = ray_o.reshape(-1, 3).astype(np.float32)
ray_d = ray_d.reshape(-1, 3).astype(np.float32)
near, far, mask_at_box = get_near_far(bounds, ray_o, ray_d)
near = near.astype(np.float32)
far = far.astype(np.float32)
rgb = rgb[mask_at_box]
ray_o = ray_o[mask_at_box]
ray_d = ray_d[mask_at_box]
coord = np.zeros([len(rgb), 2]).astype(np.int64)
return rgb, ray_o, ray_d, near, far, coord, mask_at_box
def sample_ray_h36m(img, msk, K, R, T, bounds, nrays, split):
H, W = img.shape[:2]
ray_o, ray_d = get_rays(H, W, K, R, T)
pose = np.concatenate([R, T], axis=1)
bound_mask = get_bound_2d_mask(bounds, K, pose, H, W)
msk = msk * bound_mask
bound_mask[msk == 100] = 0
if split == 'train':
nsampled_rays = 0
face_sample_ratio = cfg.face_sample_ratio
body_sample_ratio = cfg.body_sample_ratio
ray_o_list = []
ray_d_list = []
rgb_list = []
near_list = []
far_list = []
coord_list = []
mask_at_box_list = []
while nsampled_rays < nrays:
n_body = int((nrays - nsampled_rays) * body_sample_ratio)
n_face = int((nrays - nsampled_rays) * face_sample_ratio)
n_rand = (nrays - nsampled_rays) - n_body - n_face
# sample rays on body
coord_body = np.argwhere(msk == 1)
coord_body = coord_body[np.random.randint(0, len(coord_body),
n_body)]
# sample rays on face
coord_face = np.argwhere(msk == 13)
if len(coord_face) > 0:
coord_face = coord_face[np.random.randint(
0, len(coord_face), n_face)]
# sample rays in the bound mask
coord = np.argwhere(bound_mask == 1)
coord = coord[np.random.randint(0, len(coord), n_rand)]
if len(coord_face) > 0:
coord = np.concatenate([coord_body, coord_face, coord], axis=0)
else:
coord = np.concatenate([coord_body, coord], axis=0)
ray_o_ = ray_o[coord[:, 0], coord[:, 1]]
ray_d_ = ray_d[coord[:, 0], coord[:, 1]]
rgb_ = img[coord[:, 0], coord[:, 1]]
near_, far_, mask_at_box = get_near_far(bounds, ray_o_, ray_d_)
ray_o_list.append(ray_o_[mask_at_box])
ray_d_list.append(ray_d_[mask_at_box])
rgb_list.append(rgb_[mask_at_box])
near_list.append(near_)
far_list.append(far_)
coord_list.append(coord[mask_at_box])
mask_at_box_list.append(mask_at_box[mask_at_box])
nsampled_rays += len(near_)
ray_o = np.concatenate(ray_o_list).astype(np.float32)
ray_d = np.concatenate(ray_d_list).astype(np.float32)
rgb = np.concatenate(rgb_list).astype(np.float32)
near = np.concatenate(near_list).astype(np.float32)
far = np.concatenate(far_list).astype(np.float32)
coord = np.concatenate(coord_list)
mask_at_box = np.concatenate(mask_at_box_list)
else:
rgb = img.reshape(-1, 3).astype(np.float32)
ray_o = ray_o.reshape(-1, 3).astype(np.float32)
ray_d = ray_d.reshape(-1, 3).astype(np.float32)
near, far, mask_at_box = get_near_far(bounds, ray_o, ray_d)
near = near.astype(np.float32)
far = far.astype(np.float32)
rgb = rgb[mask_at_box]
ray_o = ray_o[mask_at_box]
ray_d = ray_d[mask_at_box]
coord = np.zeros([len(rgb), 2]).astype(np.int64)
return rgb, ray_o, ray_d, near, far, coord, mask_at_box
def get_smpl_data(ply_path):
ply = trimesh.load(ply_path)
xyz = np.array(ply.vertices)
nxyz = np.array(ply.vertex_normals)
if cfg.add_pointcloud:
# add random points
xyz_, ind_ = trimesh.sample.sample_surface_even(ply, 5000)
nxyz_ = ply.face_normals[ind_]
xyz = np.concatenate([xyz, xyz_], axis=0)
nxyz = np.concatenate([nxyz, nxyz_], axis=0)
xyz = xyz.astype(np.float32)
nxyz = nxyz.astype(np.float32)
return xyz, nxyz
def get_acc(coord, msk):
border = 25
kernel = np.ones((border, border), np.uint8)
msk = cv2.dilate(msk.copy(), kernel)
acc = msk[coord[:, 0], coord[:, 1]]
acc = (acc != 0).astype(np.uint8)
return acc
def rotate_smpl(xyz, nxyz, t):
"""
t: rotation angle
"""
xyz = xyz.copy()
nxyz = nxyz.copy()
center = (np.min(xyz, axis=0) + np.max(xyz, axis=0)) / 2
xyz = xyz - center
R = np.array([[np.cos(t), -np.sin(t)], [np.sin(t), np.cos(t)]])
R = R.astype(np.float32)
xyz[:, :2] = np.dot(xyz[:, :2], R.T)
xyz = xyz + center
# nxyz[:, :2] = np.dot(nxyz[:, :2], R.T)
return xyz, nxyz, center
def transform_can_smpl(xyz):
center = np.array([0, 0, 0]).astype(np.float32)
rot = np.array([[np.cos(0), -np.sin(0)], [np.sin(0), np.cos(0)]])
rot = rot.astype(np.float32)
trans = np.array([0, 0, 0]).astype(np.float32)
if np.random.uniform() > cfg.rot_ratio:
return xyz, center, rot, trans
xyz = xyz.copy()
# rotate the smpl
rot_range = np.pi / 32
t = np.random.uniform(-rot_range, rot_range)
rot = np.array([[np.cos(t), -np.sin(t)], [np.sin(t), np.cos(t)]])
rot = rot.astype(np.float32)
center = np.mean(xyz, axis=0)
xyz = xyz - center
xyz[:, [0, 2]] = np.dot(xyz[:, [0, 2]], rot.T)
xyz = xyz + center
# translate the smpl
x_range = 0.05
z_range = 0.025
x_trans = np.random.uniform(-x_range, x_range)
z_trans = np.random.uniform(-z_range, z_range)
trans = np.array([x_trans, 0, z_trans]).astype(np.float32)
xyz = xyz + trans
return xyz, center, rot, trans
def unproject(depth, K, R, T):
H, W = depth.shape
i, j = np.meshgrid(np.arange(W, dtype=np.float32),
np.arange(H, dtype=np.float32),
indexing='xy')
xy1 = np.stack([i, j, np.ones_like(i)], axis=2)
xyz = xy1 * depth[..., None]
pts3d = np.dot(xyz, np.linalg.inv(K).T)
pts3d = np.dot(pts3d - T.ravel(), R)
return pts3d
def sample_world_points(ray_o, ray_d, near, far, split):
# calculate the steps for each ray
t_vals = np.linspace(0., 1., num=cfg.N_samples)
z_vals = near[..., None] * (1. - t_vals) + far[..., None] * t_vals
if cfg.perturb > 0. and split == 'train':
# get intervals between samples
mids = .5 * (z_vals[..., 1:] + z_vals[..., :-1])
upper = np.concatenate([mids, z_vals[..., -1:]], -1)
lower = np.concatenate([z_vals[..., :1], mids], -1)
# stratified samples in those intervals
t_rand = np.random.rand(*z_vals.shape)
z_vals = lower + (upper - lower) * t_rand
pts = ray_o[:, None] + ray_d[:, None] * z_vals[..., None]
pts = pts.astype(np.float32)
z_vals = z_vals.astype(np.float32)
return pts, z_vals
def barycentric_interpolation(val, coords):
"""
:param val: verts x 3 x d input matrix
:param coords: verts x 3 barycentric weights array
:return: verts x d weighted matrix
"""
t = val * coords[..., np.newaxis]
ret = t.sum(axis=1)
return ret
def batch_rodrigues(poses):
""" poses: N x 3
"""
batch_size = poses.shape[0]
angle = np.linalg.norm(poses + 1e-8, axis=1, keepdims=True)
rot_dir = poses / angle
cos = np.cos(angle)[:, None]
sin = np.sin(angle)[:, None]
rx, ry, rz = np.split(rot_dir, 3, axis=1)
zeros = np.zeros([batch_size, 1])
K = np.concatenate([zeros, -rz, ry, rz, zeros, -rx, -ry, rx, zeros], axis=1)
K = K.reshape([batch_size, 3, 3])
ident = np.eye(3)[None]
rot_mat = ident + sin * K + (1 - cos) * np.matmul(K, K)
return rot_mat
def get_rigid_transformation(poses, joints, parents):
"""
poses: 24 x 3
joints: 24 x 3
parents: 24
"""
rot_mats = batch_rodrigues(poses)
# obtain the relative joints
rel_joints = joints.copy()
rel_joints[1:] -= joints[parents[1:]]
# create the transformation matrix
transforms_mat = np.concatenate([rot_mats, rel_joints[..., None]], axis=2)
padding = np.zeros([24, 1, 4])
padding[..., 3] = 1
transforms_mat = np.concatenate([transforms_mat, padding], axis=1)
# rotate each part
transform_chain = [transforms_mat[0]]
for i in range(1, parents.shape[0]):
curr_res = np.dot(transform_chain[parents[i]], transforms_mat[i])
transform_chain.append(curr_res)
transforms = np.stack(transform_chain, axis=0)
# obtain the rigid transformation
padding = np.zeros([24, 1])
joints_homogen = np.concatenate([joints, padding], axis=1)
transformed_joints = np.sum(transforms * joints_homogen[:, None], axis=2)
transforms[..., 3] = transforms[..., 3] - transformed_joints
transforms = transforms.astype(np.float32)
return transforms
|