Spaces:
Build error
Build error
File size: 12,309 Bytes
d7a991a |
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 |
# Copyright (c) OpenMMLab. All rights reserved.
import argparse
import os
import pickle
import shutil
from os.path import join
import cv2
import h5py
import mmcv
import numpy as np
from scipy.io import loadmat
train_subjects = [i for i in range(1, 9)]
test_subjects = [i for i in range(1, 7)]
train_seqs = [1, 2]
train_cams = [0, 1, 2, 4, 5, 6, 7, 8]
train_frame_nums = {
(1, 1): 6416,
(1, 2): 12430,
(2, 1): 6502,
(2, 2): 6081,
(3, 1): 12488,
(3, 2): 12283,
(4, 1): 6171,
(4, 2): 6675,
(5, 1): 12820,
(5, 2): 12312,
(6, 1): 6188,
(6, 2): 6145,
(7, 1): 6239,
(7, 2): 6320,
(8, 1): 6468,
(8, 2): 6054
}
test_frame_nums = {1: 6151, 2: 6080, 3: 5838, 4: 6007, 5: 320, 6: 492}
train_img_size = (2048, 2048)
root_index = 14
joints_17 = [7, 5, 14, 15, 16, 9, 10, 11, 23, 24, 25, 18, 19, 20, 4, 3, 6]
def get_pose_stats(kps):
"""Get statistic information `mean` and `std` of pose data.
Args:
kps (ndarray): keypoints in shape [..., K, C] where K and C is
the keypoint category number and dimension.
Returns:
mean (ndarray): [K, C]
"""
assert kps.ndim > 2
K, C = kps.shape[-2:]
kps = kps.reshape(-1, K, C)
mean = kps.mean(axis=0)
std = kps.std(axis=0)
return mean, std
def get_annotations(joints_2d, joints_3d, scale_factor=1.2):
"""Get annotations, including centers, scales, joints_2d and joints_3d.
Args:
joints_2d: 2D joint coordinates in shape [N, K, 2], where N is the
frame number, K is the joint number.
joints_3d: 3D joint coordinates in shape [N, K, 3], where N is the
frame number, K is the joint number.
scale_factor: Scale factor of bounding box. Default: 1.2.
Returns:
centers (ndarray): [N, 2]
scales (ndarray): [N,]
joints_2d (ndarray): [N, K, 3]
joints_3d (ndarray): [N, K, 4]
"""
# calculate joint visibility
visibility = (joints_2d[:, :, 0] >= 0) * \
(joints_2d[:, :, 0] < train_img_size[0]) * \
(joints_2d[:, :, 1] >= 0) * \
(joints_2d[:, :, 1] < train_img_size[1])
visibility = np.array(visibility, dtype=np.float32)[:, :, None]
joints_2d = np.concatenate([joints_2d, visibility], axis=-1)
joints_3d = np.concatenate([joints_3d, visibility], axis=-1)
# calculate bounding boxes
bboxes = np.stack([
np.min(joints_2d[:, :, 0], axis=1),
np.min(joints_2d[:, :, 1], axis=1),
np.max(joints_2d[:, :, 0], axis=1),
np.max(joints_2d[:, :, 1], axis=1)
],
axis=1)
centers = np.stack([(bboxes[:, 0] + bboxes[:, 2]) / 2,
(bboxes[:, 1] + bboxes[:, 3]) / 2],
axis=1)
scales = scale_factor * np.max(bboxes[:, 2:] - bboxes[:, :2], axis=1) / 200
return centers, scales, joints_2d, joints_3d
def load_trainset(data_root, out_dir):
"""Load training data, create annotation file and camera file.
Args:
data_root: Directory of dataset, which is organized in the following
hierarchy:
data_root
|-- train
|-- S1
|-- Seq1
|-- Seq2
|-- S2
|-- ...
|-- test
|-- TS1
|-- TS2
|-- ...
out_dir: Directory to save annotation file.
"""
_imgnames = []
_centers = []
_scales = []
_joints_2d = []
_joints_3d = []
cameras = {}
img_dir = join(out_dir, 'images')
os.makedirs(img_dir, exist_ok=True)
annot_dir = join(out_dir, 'annotations')
os.makedirs(annot_dir, exist_ok=True)
for subj in train_subjects:
for seq in train_seqs:
seq_path = join(data_root, 'train', f'S{subj}', f'Seq{seq}')
num_frames = train_frame_nums[(subj, seq)]
# load camera parametres
camera_file = join(seq_path, 'camera.calibration')
with open(camera_file, 'r') as fin:
lines = fin.readlines()
for cam in train_cams:
K = [float(s) for s in lines[cam * 7 + 5][11:-2].split()]
f = np.array([[K[0]], [K[5]]])
c = np.array([[K[2]], [K[6]]])
RT = np.array(
[float(s) for s in lines[cam * 7 + 6][11:-2].split()])
RT = np.reshape(RT, (4, 4))
R = RT[:3, :3]
# convert unit from millimeter to meter
T = RT[:3, 3:] * 0.001
size = [int(s) for s in lines[cam * 7 + 3][14:].split()]
w, h = size
cam_param = dict(
R=R, T=T, c=c, f=f, w=w, h=h, name=f'train_cam_{cam}')
cameras[f'S{subj}_Seq{seq}_Cam{cam}'] = cam_param
# load annotations
annot_file = os.path.join(seq_path, 'annot.mat')
annot2 = loadmat(annot_file)['annot2']
annot3 = loadmat(annot_file)['annot3']
for cam in train_cams:
# load 2D and 3D annotations
joints_2d = np.reshape(annot2[cam][0][:num_frames],
(num_frames, 28, 2))[:, joints_17]
joints_3d = np.reshape(annot3[cam][0][:num_frames],
(num_frames, 28, 3))[:, joints_17]
joints_3d = joints_3d * 0.001
centers, scales, joints_2d, joints_3d = get_annotations(
joints_2d, joints_3d)
_centers.append(centers)
_scales.append(scales)
_joints_2d.append(joints_2d)
_joints_3d.append(joints_3d)
# extract frames from video
video_path = join(seq_path, 'imageSequence',
f'video_{cam}.avi')
video = mmcv.VideoReader(video_path)
for i in mmcv.track_iter_progress(range(num_frames)):
img = video.read()
if img is None:
break
imgname = f'S{subj}_Seq{seq}_Cam{cam}_{i+1:06d}.jpg'
_imgnames.append(imgname)
cv2.imwrite(join(img_dir, imgname), img)
_imgnames = np.array(_imgnames)
_centers = np.concatenate(_centers)
_scales = np.concatenate(_scales)
_joints_2d = np.concatenate(_joints_2d)
_joints_3d = np.concatenate(_joints_3d)
out_file = join(annot_dir, 'mpi_inf_3dhp_train.npz')
np.savez(
out_file,
imgname=_imgnames,
center=_centers,
scale=_scales,
part=_joints_2d,
S=_joints_3d)
print(f'Create annotation file for trainset: {out_file}. '
f'{len(_imgnames)} samples in total.')
out_file = join(annot_dir, 'cameras_train.pkl')
with open(out_file, 'wb') as fout:
pickle.dump(cameras, fout)
print(f'Create camera file for trainset: {out_file}.')
# get `mean` and `std` of pose data
_joints_3d = _joints_3d[..., :3] # remove visibility
mean_3d, std_3d = get_pose_stats(_joints_3d)
_joints_2d = _joints_2d[..., :2] # remove visibility
mean_2d, std_2d = get_pose_stats(_joints_2d)
# centered around root
_joints_3d_rel = _joints_3d - _joints_3d[..., root_index:root_index + 1, :]
mean_3d_rel, std_3d_rel = get_pose_stats(_joints_3d_rel)
mean_3d_rel[root_index] = mean_3d[root_index]
std_3d_rel[root_index] = std_3d[root_index]
_joints_2d_rel = _joints_2d - _joints_2d[..., root_index:root_index + 1, :]
mean_2d_rel, std_2d_rel = get_pose_stats(_joints_2d_rel)
mean_2d_rel[root_index] = mean_2d[root_index]
std_2d_rel[root_index] = std_2d[root_index]
stats = {
'joint3d_stats': {
'mean': mean_3d,
'std': std_3d
},
'joint2d_stats': {
'mean': mean_2d,
'std': std_2d
},
'joint3d_rel_stats': {
'mean': mean_3d_rel,
'std': std_3d_rel
},
'joint2d_rel_stats': {
'mean': mean_2d_rel,
'std': std_2d_rel
}
}
for name, stat_dict in stats.items():
out_file = join(annot_dir, f'{name}.pkl')
with open(out_file, 'wb') as f:
pickle.dump(stat_dict, f)
print(f'Create statistic data file: {out_file}')
def load_testset(data_root, out_dir, valid_only=True):
"""Load testing data, create annotation file and camera file.
Args:
data_root: Directory of dataset.
out_dir: Directory to save annotation file.
valid_only: Only keep frames with valid_label == 1.
"""
_imgnames = []
_centers = []
_scales = []
_joints_2d = []
_joints_3d = []
cameras = {}
img_dir = join(out_dir, 'images')
os.makedirs(img_dir, exist_ok=True)
annot_dir = join(out_dir, 'annotations')
os.makedirs(annot_dir, exist_ok=True)
for subj in test_subjects:
subj_path = join(data_root, 'test', f'TS{subj}')
num_frames = test_frame_nums[subj]
# load annotations
annot_file = os.path.join(subj_path, 'annot_data.mat')
with h5py.File(annot_file, 'r') as fin:
annot2 = np.array(fin['annot2']).reshape((-1, 17, 2))
annot3 = np.array(fin['annot3']).reshape((-1, 17, 3))
valid = np.array(fin['valid_frame']).reshape(-1)
# manually estimate camera intrinsics
fx, cx = np.linalg.lstsq(
annot3[:, :, [0, 2]].reshape((-1, 2)),
(annot2[:, :, 0] * annot3[:, :, 2]).reshape(-1, 1),
rcond=None)[0].flatten()
fy, cy = np.linalg.lstsq(
annot3[:, :, [1, 2]].reshape((-1, 2)),
(annot2[:, :, 1] * annot3[:, :, 2]).reshape(-1, 1),
rcond=None)[0].flatten()
if subj <= 4:
w, h = 2048, 2048
else:
w, h = 1920, 1080
cameras[f'TS{subj}'] = dict(
c=np.array([[cx], [cy]]),
f=np.array([[fx], [fy]]),
w=w,
h=h,
name=f'test_cam_{subj}')
# get annotations
if valid_only:
valid_frames = np.nonzero(valid)[0]
else:
valid_frames = np.arange(num_frames)
joints_2d = annot2[valid_frames, :, :]
joints_3d = annot3[valid_frames, :, :] * 0.001
centers, scales, joints_2d, joints_3d = get_annotations(
joints_2d, joints_3d)
_centers.append(centers)
_scales.append(scales)
_joints_2d.append(joints_2d)
_joints_3d.append(joints_3d)
# copy and rename images
for i in valid_frames:
imgname = f'TS{subj}_{i+1:06d}.jpg'
shutil.copyfile(
join(subj_path, 'imageSequence', f'img_{i+1:06d}.jpg'),
join(img_dir, imgname))
_imgnames.append(imgname)
_imgnames = np.array(_imgnames)
_centers = np.concatenate(_centers)
_scales = np.concatenate(_scales)
_joints_2d = np.concatenate(_joints_2d)
_joints_3d = np.concatenate(_joints_3d)
if valid_only:
out_file = join(annot_dir, 'mpi_inf_3dhp_test_valid.npz')
else:
out_file = join(annot_dir, 'mpi_inf_3dhp_test_all.npz')
np.savez(
out_file,
imgname=_imgnames,
center=_centers,
scale=_scales,
part=_joints_2d,
S=_joints_3d)
print(f'Create annotation file for testset: {out_file}. '
f'{len(_imgnames)} samples in total.')
out_file = join(annot_dir, 'cameras_test.pkl')
with open(out_file, 'wb') as fout:
pickle.dump(cameras, fout)
print(f'Create camera file for testset: {out_file}.')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('data_root', type=str, help='data root')
parser.add_argument(
'out_dir', type=str, help='directory to save annotation files.')
args = parser.parse_args()
data_root = args.data_root
out_dir = args.out_dir
load_trainset(data_root, out_dir)
load_testset(data_root, out_dir, valid_only=True)
|