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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454
| import numpy as np import os,imageio from subprocess import check_output
def render_path_spiral(c2w, up, rads, focal, zdelta, zrate, rots, N): render_poses = [] rads = np.array(list(rads) + [1.]) hwf = c2w[:, 4:5] """ c是当前迭代的相机在世界坐标系的位置, np.dot(c2w[:3, :4], np.array([0, 0, -focal, 1.])是焦点在世界坐标系的位置 z是相机z轴在世界坐标系的朝向。 normalize_c2w_matrix(z, up, c)构造当前相机的参数矩阵。 """
for theta in np.linspace(0., 2. * np.pi * rots, N + 1)[:-1]: c = np.dot(c2w[:3, :4], np.array([np.cos(theta), -np.sin(theta), -np.sin(theta * zrate), 1.]) * rads) z = normalize(c - np.dot(c2w[:3, :4], np.array([0, 0, -focal, 1.]))) render_poses.append(np.concatenate([normalize_c2w_matrix(z, up, c), hwf], 1)) return render_poses
def spherify_poses(poses, bds): p34_to_44 = lambda p: np.concatenate([p, np.tile(np.reshape(np.eye(4)[-1, :], [1, 1, 4]), [p.shape[0], 1, 1])], 1)
rays_d = poses[:, :3, 2:3] rays_o = poses[:, :3, 3:4]
def min_line_dist(rays_o, rays_d): A_i = np.eye(3) - rays_d * np.transpose(rays_d, [0, 2, 1]) b_i = -A_i @ rays_o pt_mindist = np.squeeze(-np.linalg.inv((np.transpose(A_i, [0, 2, 1]) @ A_i).mean(0)) @ (b_i).mean(0)) return pt_mindist
pt_mindist = min_line_dist(rays_o, rays_d)
center = pt_mindist up = (poses[:, :3, 3] - center).mean(0)
vec0 = normalize(up) vec1 = normalize(np.cross([.1, .2, .3], vec0)) vec2 = normalize(np.cross(vec0, vec1)) pos = center c2w = np.stack([vec1, vec2, vec0, pos], 1)
poses_reset = np.linalg.inv(p34_to_44(c2w[None])) @ p34_to_44(poses[:, :3, :4])
rad = np.sqrt(np.mean(np.sum(np.square(poses_reset[:, :3, 3]), -1)))
sc = 1. / rad poses_reset[:, :3, 3] *= sc bds *= sc rad *= sc
centroid = np.mean(poses_reset[:, :3, 3], 0) zh = centroid[2] radcircle = np.sqrt(rad ** 2 - zh ** 2) new_poses = []
for th in np.linspace(0., 2. * np.pi, 120): camorigin = np.array([radcircle * np.cos(th), radcircle * np.sin(th), zh]) up = np.array([0, 0, -1.])
vec2 = normalize(camorigin) vec0 = normalize(np.cross(vec2, up)) vec1 = normalize(np.cross(vec2, vec0)) pos = camorigin p = np.stack([vec0, vec1, vec2, pos], 1)
new_poses.append(p)
new_poses = np.stack(new_poses, 0)
new_poses = np.concatenate([new_poses, np.broadcast_to(poses[0, :3, -1:], new_poses[:, :3, -1:].shape)], -1) poses_reset = np.concatenate( [poses_reset[:, :3, :4], np.broadcast_to(poses[0, :3, -1:], poses_reset[:, :3, -1:].shape)], -1)
return poses_reset, new_poses, bds
def normalize_c2w_matrix(norm_z,norm_y,mean_center):
""" 原函数 里面有点看不懂 def viewmatrix(z, up, pos): vec2 = normalize(z) vec1_avg = up vec0 = normalize(np.cross(vec1_avg, vec2)) vec1 = normalize(np.cross(vec2, vec0)) m = np.stack([vec0, vec1, vec2, pos], 1) return m """ norm_vec_z = nomailize(norm_z) norm_vec_y = norm_y norm_vec_x = normalize(np.cross(norm_vec_y,norm_vec_z)) norm_vec_y = normalize(np.cross(norm_vec_z,norm_vec_x)) c2w = np.stack([norm_vec_x,norm_vec_y,norm_z,mean_center],1) return c2w
def normalize(x): return x/np.linalg.norm(x)
def poses_average(poses): """
:param poses: [vec_X,vec_Y,vec_Z,center,hwf] 也叫C2W矩阵 :return: c2w camera to world """
hwf = poses[0,:3,-1:] mean_center = poses[:,:3,3].mean(0) norm_z = normalize(poses[:,:3,2].sum(0)) norm_y = poses[:,:3,1].sum(0) poses = np.concatenate([normalize_c2w_matrix(norm_z,norm_y,mean_center),hwf],1) return poses
def recenter_poses(poses): """
:param poses: 相机所有的位姿数据 :return: """ """ 首先我们要知道利用同一个旋转平移变换矩阵左乘所有的相机位姿是对所有的相机位姿做一个全局的旋转平移变换 我们可以用平均相机位姿作为支点理解, 如果把平均位姿的逆c2w^-1左乘平均相机位姿pose_avg,返回的相机位姿中旋转矩阵为单位矩阵,平移量为零向量。 也就是变换后的平均相机位姿的位置处在世界坐标系的原点,XYZ轴朝向和世界坐标系的向一致。 """ poses_ = poses+0 bottom = np.reshape([0,0,0,1.0],[1,4]) poses_avg = poses_average(poses) c2w_avg = np.concatenate([poses_avg[:3,:4],bottom],0) bottom = np.tile(np.reshape(bottom, [1, 1, 4]), [poses.shape[0], 1, 1]) poses = np.concatenate([poses[:,:3,:4],bottom],-2) poses = np.linalg.inv(c2w_avg) @ poses poses_[:, :3, :4] = poses[:, :3, :4] poses = poses_ return poses
def imread(file): if file.endswith("png"): return imageio.imread(file,ignoregamma=True) else: return imageio.imread(file)
def _minify(datadir,factors=[],resolutions=[]): needtoload = False
for factor in factors: imgdir = os.path.join(datadir,"images_{}".format(factor)) if not os.path.exists(imgdir): needtoload = True for resolution in resolutions: imgdir = os.path.join(datadir,"images_{}x{}".format(resolution[1],resolution[0])) if not os.path.exists(imgdir): needtoload = True
if not needtoload: return
imgdir = os.path.join(datadir,"images") imgs = [os.path.join(imgdir,file) for file in sorted(os.listdir(imgdir))] imgs = [file for file in imgs if any([file.endswith(ex) for ex in ["JPG","jpg","PNG","JPEG","PNG"]])]
imgdir_original = imgdir
img_workDir = os.getcwd()
for r in factors+resolutions: if isinstance(r,int): name = "images_{}".format(r) resize_argument = "{}%".format(100./r) else: name = "images_{}x{}".format(r[1],r[0]) resize_argument = "{}x{}".format(r[1],r[0])
imgdir = os.path.join(datadir,name) if os.path.exists(imgdir): continue
print("Minifying",r,datadir)
os.makedirs(imgdir) check_output('cp {}/*{}'.format(imgdir_original,imgdir),shell=True) img_extension = imgs[0].split('.')[-1]
shell_args = ' '.join(['mogrify','-resize',resize_argument,'-format','png','*.{}'.format(img_extension)]) print(shell_args) os.chdir() os.chdir(imgdir) check_output(shell_args,shell=True) os.chdir(img_workDir)
if img_extension != 'png': check_output('rm {}/*.{}'.format(imgdir,img_extension),shell=True) print("removed duplicates which are not png") print("Done")
def get_poses_bds_imgs(datadir, factor=None, width=None, height=None, load_imgs = True):
""" # 20 x 17 位姿前15个 near far 后两个 """
""" poses_bounds.npy 文件 [ r11 r12 r13 t1 H r21 r22 r23 t2 W 3 x 5 在npy文件里压缩成 --> 1 x 15 [r11 r12 r13 t1 H r21 r22 r23 t2 W r31 r32 r33 t3 f] r31 r32 r33 t3 f]
相机外参的逆矩阵被称为camera-to-world (c2w)矩阵,其作用是把相机坐标系的点变换到世界坐标系。 r11~r33 c2w旋转矩阵 t1~t3 c2w平移向量 H 照片高度 W 照片宽度 f 相机焦距 最后两个参数用于表示场景的范围Bounds (bds),是该相机视角下场景点离相机中心最近(near)和最远(far)的距离。 [r11 r12 r13 t1 H r21 r22 r23 t2 W r31 r32 r33 t3 f near far] """ poses_mat = np.load(os.path.join(datadir,'poses_bounds.npy')) c2w = poses_mat[:,:-2].reshape([-1,3,5]).transpose([1,2,0]) bds = poses_mat[:,-2:].transpose([1,0])
img0 = [os.path.join(datadir,'images',f) for f in sorted(os.listdir(os.path.join(datadir,'images'))) if f.endswith("JPG") or f.endswith("jpg") or f.endswith("png")][0] raw_imageshape = imageio.imread(img0).shape
sfx = '' if factor is not None: sfx ="_{}".format(factor) _minify(datadir,factors=[factor]) factor = factor elif height is not None: factor = raw_imageshape[0]/float[height] width = int(raw_imageshape[1]/factor) _minify(datadir,resolutions=[[height,width]]) sfx = "_{}x{}".format(width,height) elif width is not None: factor = raw_imageshape[1]/float[width] height = int(raw_imageshape[0]/factor) _minify(datadir,resolutions=[[height,width]]) sfx = "_{}x{}".format(width,height) else: factor = 1
imgdir = os.path.join(datadir,'images'+sfx) if not os.path.exists(imgdir): print(imgdir,"does not exists,returning") return
imgfiles =[os.path.join(imgdir,file) for file in sorted(os.listdir(imgdir)) if file.endswith("JPG") or file.endswith("jpg") or file.endswith("png")]
if c2w.shape[-1] !=len(imgfiles): print("Mismatching between img{} and poses {} !!!! ".format((len(imgfiles)),c2w.shape[-1])) return
shx_imageshape = imageio.imread(imgfiles[0]).shape
poses = c2w poses[:2,4,:] = np.array(shx_imageshape[:2]).reshape([2,1]) poses[2,4,:] = c2w[2,4,:]*1.0/factor
if not load_imgs: return poses,bds
imgs = [imread(file)[...,:3]/255 for file in imgfiles] imgs = np.stack(imgs,-1)
print("load image data",imgs.shape,poses[:,-1,0]) return poses,bds,imgs
def load_llff_data( datadir, factor=8, recenter=True, bd_factor=0.75, spherify = False, path_zflat=False): poses,bds,imgs = get_poses_bds_imgs(datadir,factor) print("Loaded",datadir,bds.min(),bds.max())
poses = np.concatenate([poses[:,1:2,:],-poses[:,0:1,:],poses[:,2:,:]],1) poses = np.moveaxis(poses,-1,0).astype(np.float32) imgs = np.moveaxis(imgs,-1,0).astype(np.float32) images = imgs bds = np.moveaxis(bds,-1,0).astype(np.float32)
sc = 1.0 if bd_factor is None else 1.0/(bds.min() *bd_factor) poses[:,:3,3] *= sc bds *= sc
if recenter: poses = recenter_poses(poses)
if spherify: poses, render_poses,bds = spherify_poses(poses,bds)
else:
""" 准备参数 """ c2w = poses_average(poses) print("recentered",c2w.shape) print(c2w[:3,:4])
up = normalize(poses[:,:3,1].sum(0))
close_depth,inf_depth = bds.min()*0.9,bds.max()*5.0 dt = 0.75 mean_depth = 1.0/(((1.0-dt)/close_depth+dt/inf_depth)) focal = mean_depth
shrink_factor = .8 zdelta = close_depth * .2 tt = poses[:, :3, 3] rads = np.percentile(np.abs(tt), 90, 0) c2w_path = c2w N_views = 120 N_rots = 2 if path_zflat: zloc = -close_depth * .1 c2w_path[:3, 3] = c2w_path[:3, 3] + zloc * c2w_path[:3, 2] rads[2] = 0. N_rots = 1 N_views /= 2
render_poses = render_path_spiral(c2w_path, up, rads, focal, zdelta, zrate=.5, rots=N_rots, N=N_views)
render_poses = np.array(render_poses).astype(np.float32) c2w = poses_average(poses) print('Data:') print(poses.shape, images.shape, bds.shape)
dists = np.sum(np.square(c2w[:3, 3] - poses[:, :3, 3]), -1) i_test = np.argmin(dists) print('HOLDOUT view is', i_test)
images = images.astype(np.float32) poses = poses.astype(np.float32)
return images, poses, bds, render_poses, i_test
|