def prepare_data(self):
for split_name, value in self.split_dict.items():
image_paths, kps_2d, kps_3d, sequences = [], [], [], []
sub_ids = value['sub_ids']
seq_ids = value['seq_ids']
cam_ids = value['cam_ids']
idx = 0
for sub_id in sub_ids:
print('convert subject {} {}/{}'.format(sub_id, sub_ids.index(sub_id) + 1, len(sub_ids)))
if seq_ids is None and cam_ids is None:
img_dir, image_ids, kp2d, kp3d = self.convert_test(sub_id)
for i, image_id in enumerate(image_ids):
image_paths.append(img_dir % image_id)
kps_2d.append(kp2d[i])
kps_3d.append(kp3d[i])
sequences.append('mpii_ts{}'.format(sub_id))
else:
for seq_id in seq_ids:
for cam_id in cam_ids:
img_dir, num_images, kp2d, kp3d = self.convert_train(sub_id, seq_id, cam_id)
idx += 1
for i in range(num_images):
image_paths.append(img_dir % (i + 1))
kps_2d.append(kp2d[i])
kps_3d.append(kp3d[i])
image_paths = np.asarray(image_paths)
kps_2d = np.asarray(kps_2d)
kps_3d = np.asarray(kps_3d)
sequences = np.asarray(sequences) if split_name == 'test' else None
mpii_config = DataSetConfig(split_name, True, self.mpii_3d_order, lsp_only=True)
self.data_set_splits.append(DataSetSplit(mpii_config, image_paths, kps_2d, kps_3d=kps_3d, seqs=sequences))
def create_dataset_split(self, annotations, img_ids, name):
def convert_vis(value):
if type(value) == int or (type(value) == str
and value in ['1', '0']):
return int(value)
elif isinstance(value, np.ndarray):
return int(value.size != 0)
else:
return 0
images, kps_2d, vis = [], [], []
img_dir = join(self.data_dir, 'images')
print("prepare {} mpii annotations for conversion".format(name))
for img_id in tqdm(img_ids):
try:
ann_info = annotations.annolist[img_id]
single_persons = annotations.single_person[img_id]
if not isinstance(single_persons, np.ndarray):
single_persons = np.array([single_persons])
if single_persons.size == 0:
continue
rects = ann_info.annorect
if not isinstance(rects, np.ndarray):
rects = np.array([rects])
persons = rects[single_persons - 1]
for person in persons:
points = person.annopoints.point
if not isinstance(points, np.ndarray):
# There is only one! so ignore this image
continue
kp2d = np.zeros((self.num_kps, 2), np.float32)
v = np.zeros((self.num_kps, ), np.float32)
for p in points:
kp2d[p.id] = [p.x, p.y]
v[p.id] = convert_vis(p.is_visible)
images.append(join(img_dir, ann_info.image.name))
kps_2d.append(kp2d)
vis.append(v)
except (AttributeError, TypeError):
print('error')
continue
images = np.asarray(images)
kps_2d = np.asarray(kps_2d, dtype=np.float32)
vis = np.asarray(vis, dtype=np.int64)
mpii_config = DataSetConfig(name, False, self.mpii_order)
return DataSetSplit(mpii_config, images, kps_2d, vis)
def prepare_data(self):
image_paths = np.array(
sorted([f for f in glob(join(self.data_dir, 'images/*.jpg'))]))
joints = sio.loadmat(join(self.data_dir, 'joints.mat'),
squeeze_me=True)['joints'].astype(np.float32)
if self.args.dataset_name == 'lsp_ext':
joints = np.transpose(joints, (2, 0, 1))
kps_2d = joints[:, :, :2]
vis = joints[:, :, 2].astype(np.int64)
lsp_config = DataSetConfig('train',
has_3d=False,
reorder=self.lsp_order)
self.data_set_splits.append(
DataSetSplit(lsp_config, image_paths, kps_2d, vis))
elif self.args.dataset_name == 'lsp':
joints = np.transpose(joints, (2, 1, 0))
kps_2d = joints[:, :, :2]
vis = (1 - joints[:, :, 2]).astype(np.int64)
lsp_config_train = DataSetConfig('train',
has_3d=False,
reorder=self.lsp_order)
self.data_set_splits.append(
DataSetSplit(lsp_config_train, image_paths[:1000],
kps_2d[:1000], vis[:1000]))
lsp_config_test = DataSetConfig('val',
has_3d=False,
reorder=self.lsp_order)
self.data_set_splits.append(
DataSetSplit(lsp_config_test, image_paths[1000:],
kps_2d[1000:], vis[1000:]))
else:
raise Exception('unknown LSP dataset name')
def prepare_data(self):
splits = ['train', 'val']
for split in splits:
image_paths, kps_2d, vis = [], [], []
data_type = split + self.args.year
img_dir = join(self.data_dir, data_type)
# load coco annotations
self.coco = COCO(
join(self.data_dir, 'annotations',
'person_keypoints_{}.json'.format(data_type)))
# get id for category person
cat_ids = self.coco.getCatIds(catNms=['person'])
# get all image id's containing instances of person category
img_ids = self.coco.getImgIds(catIds=cat_ids)
# np.random.shuffle(img_ids)
# img_ids = img_ids[:1000]
print("prepare coco annotations for conversion")
for img_id in tqdm(img_ids):
img_path = join(img_dir, '%012d.jpg' % img_id)
ann_ids = self.coco.getAnnIds(imgIds=img_id,
catIds=cat_ids,
iscrowd=False)
img_annotations = self.coco.loadAnns(ann_ids)
for ann in img_annotations:
keypoints = np.array(ann['keypoints'])
keypoints = np.reshape(keypoints, (-1, 3))
image_paths.append(img_path)
kps_2d.append(keypoints[:, :2])
vis.append(keypoints[:, 2] == 2)
image_paths = np.asarray(image_paths)
kps_2d = np.asarray(kps_2d, dtype=np.float32)
vis = np.asarray(vis, dtype=np.int64)
# generate zero placeholders
coco_config = DataSetConfig(split, False, self.coco_order,
self.face_and_shoulder)
self.data_set_splits.append(
DataSetSplit(coco_config, image_paths, kps_2d, vis))
def prepare_data(self):
for split_name, value in self.split_dict.items():
sub_ids = value['sub_ids']
seq_ids = value['seq_ids']
image_paths, kps_2d, kps_3d, sequences = [], [], [], []
idx = 0
for sub_id in sub_ids:
print('convert subject {} {}/{}'.format(sub_id, sub_ids.index(sub_id) + 1, len(sub_ids)))
ann_path = join(self.data_dir, 'S%d' % sub_id, 'Positions_3D')
seqs = [str(s.rsplit('.', 1)[0]) for s in listdir(ann_path)]
for seq in seqs:
if seq not in seq_ids:
continue
with open(join(ann_path, seq + '.csv')) as file:
reader = csv.reader(file, delimiter=';')
content = {}
cam_info = []
for row in reader:
if row[0] == '#camerainfo':
key = 'cam' + row[1]
cam_info.append(key)
else:
key = row[0]
content[key] = row[1:]
content['#camerainfo'] = cam_info
cam_infos = {cam: CameraInfo.from_line(content[cam]) for cam in content['#camerainfo']}
num_frames = int(content['#number_of_frames'][0])
num_points = len(content['#point_names'])
num_dimens = int(content['#dimensions'][0])
shape = (num_frames, num_points, num_dimens)
data = np.empty(shape, dtype=np.float32)
for f in range(num_frames):
kp3d = np.asarray(content[str(f)]).reshape((-1, 3))
data[f] = kp3d[:num_points]
for cam, info in cam_infos.items():
kp3d = np.add(np.tensordot(data, info.R, axes=(2, 1)).reshape(shape), info.T)
kp2d = kp3d[:, :, :2] * info.f / np.expand_dims(kp3d[:, :, 2], -1) + info.o
sub_dir = 'S%d' % sub_id
scene_dir = 'TC_{}_{}_{}'.format(sub_dir, seq, cam)
img_dir = join(self.data_dir, sub_dir, 'imageFrames', scene_dir, 'frame_%06d.jpg')
idx += 1
for i in range(kp2d.shape[0]):
image_paths.append(img_dir % (i + 1))
kps_2d.append(kp2d[i])
kps_3d.append(kp3d[i])
if split_name == 'test':
sequences.append(scene_dir)
image_paths = np.asarray(image_paths)
kps_2d = np.asarray(kps_2d)
kps_3d = np.asarray(kps_3d)
sequences = np.asarray(sequences) if split_name == 'test' else None
if value['skip_frames'] > 1:
skip_frames = value['skip_frames']
image_paths = image_paths[::skip_frames]
kps_2d = kps_2d[::skip_frames]
kps_3d = kps_3d[::skip_frames]
sequences = sequences[::skip_frames] if split_name == 'test' else None
tc_config = DataSetConfig(split_name, has_3d=True, reorder=self.tc_order, lsp_only=True)
self.data_set_splits.append(DataSetSplit(tc_config, image_paths, kps_2d, kps_3d=kps_3d, seqs=sequences))