def compute_cord_warp_batch(self, pair_df):
if self._warp_skip == 'full':
batch = [np.empty([self._batch_size] + [1, 8])]
elif self._warp_skip == 'mask':
batch = [np.empty([self._batch_size] + [10, 8]),
np.empty([self._batch_size, 10] + list(self._image_size))]
else:
batch = [np.empty([self._batch_size] + [72])]
i = 0
for _, p in pair_df.iterrows():
fr = self._annotations_file.loc[p['from']]
to = self._annotations_file.loc[p['to']]
kp_array1 = pose_utils.load_pose_cords_from_strings(fr['keypoints_y'],
fr['keypoints_x'])
kp_array2 = pose_utils.load_pose_cords_from_strings(to['keypoints_y'],
to['keypoints_x'])
if self._warp_skip == 'mask':
batch[0][i] = pose_transform.affine_transforms(kp_array1, kp_array2)
batch[1][i] = pose_transform.pose_masks(kp_array2, self._image_size)
elif self._warp_skip == 'full':
batch[0][i] = pose_transform.estimate_uniform_transform(kp_array1, kp_array2)
else: #sel._warp_skip == 'stn'
batch[0][i][:36] = kp_array1.reshape((-1, ))
batch[0][i][36:] = kp_array2.reshape((-1, ))
i += 1
return batch
def compute_pose_map_batch(self, pair_df, direction):
assert direction in ['to', 'from']
batch = np.empty([self._batch_size] + list(self._image_size) + [18 if self._pose_rep_type == 'hm' else 3])
i = 0
for _, p in pair_df.iterrows():
row = self._annotations_file.loc[p[direction]]
if self._cache_pose_rep:
file_name = self._tmp_pose + p[direction] + self._pose_rep_type + '.npy'
if os.path.exists(file_name):
pose = np.load(file_name)
else:
kp_array = pose_utils.load_pose_cords_from_strings(row['keypoints_y'], row['keypoints_x'])
if self._pose_rep_type == 'hm':
pose = pose_utils.cords_to_map(kp_array, self._image_size)
else:
pose = pose_transform.make_stickman(kp_array, self._image_size)
np.save(file_name, pose)
else:
kp_array = pose_utils.load_pose_cords_from_strings(row['keypoints_y'], row['keypoints_x'])
if self._pose_rep_type == 'hm':
pose = pose_utils.cords_to_map(kp_array, self._image_size)
else:
pose = pose_transform.make_stickman(kp_array, self._image_size)
batch[i] = pose
i += 1
return batch
def compute_cord_warp_batch(self, pair_df, validation=False):
batch = [np.empty([self._batch_size] + [self.num_mask, 8]),
np.empty([self._batch_size, self.num_mask] + list(self._image_size)),
np.empty([self._batch_size] + [self.num_mask, 8]),
np.empty([self._batch_size, self.num_mask] + list(self._image_size)),
np.empty([self._batch_size] + [self.num_mask, 8]),
np.empty([self._batch_size, self.num_mask] + list(self._image_size))]
i = 0
for _, p in pair_df.iterrows():
fr = self._annotations_file.loc[p['from']]
to = self._annotations_file.loc[p['to']]
kp_array1 = pose_utils.load_pose_cords_from_strings(fr['keypoints_y'],
fr['keypoints_x'])
kp_array2 = pose_utils.load_pose_cords_from_strings(to['keypoints_y'],
to['keypoints_x'])
if validation:
npy_path_from = os.path.join(self._images_dir_test, p['from'])
npy_path_from = npy_path_from[:-3]+'npy'
npy_path_to = os.path.join(self._images_dir_test, p['to'])
npy_path_to = npy_path_to[:-3] + 'npy'
else:
npy_path_from = os.path.join(self._images_dir_train, p['from'])
npy_path_from = npy_path_from[:-3]+'npy'
npy_path_to = os.path.join(self._images_dir_train, p['to'])
npy_path_to = npy_path_to[:-3] + 'npy'
batch[0][i] = pose_transform.affine_transforms(kp_array1, kp_array2, self._image_size, self.use_body_mask)
batch[1][i] = pose_transform.pose_masks(kp_array2, self._image_size, self.use_body_mask, self.use_mask, npy_path_to, self.fat)
batch[2][i] = np.c_[np.ones([10,1]),np.zeros([10,3]),np.ones([10,1]),np.zeros([10,3])]
batch[3][i] = batch[1][i]
batch[4][i] = pose_transform.affine_transforms(kp_array2, kp_array1, self._image_size, self.use_body_mask)
batch[5][i] = pose_transform.pose_masks(kp_array1, self._image_size, self.use_body_mask, self.use_mask, npy_path_from, self.fat)
i += 1
return batch
def get_gaussian_mask(self, P2_name, img_size):
to = self.annos.loc[P2_name]
kp_array2 = load_pose_cords_from_strings(to['keypoints_y'],
to['keypoints_x'])
BP2_mask = make_gaussain_limb_masks(kp_array2, img_size) # BP2 mask
return BP2_mask
def create_masked_image(names, images, annotation_file):
import pose_utils
masked_images = []
df = pd.read_csv(annotation_file, sep=':')
for name, image in zip(names, images):
to = name[1]
ano_to = df[df['name'] == to].iloc[0]
kp_to = pose_utils.load_pose_cords_from_strings(ano_to['keypoints_y'], ano_to['keypoints_x'])
mask = pose_utils.produce_ma_mask(kp_to, image.shape[:2])
masked_images.append(image * mask[..., np.newaxis])
return masked_images
def check_valid(x):
kp_array = pose_utils.load_pose_cords_from_strings(
x['keypoints_y'], x['keypoints_x'])
distractor = x['name'].startswith('-1') or x['name'].startswith('0000')
return pose_check_valid(kp_array) and not distractor
args.annotations_file_test = 'data/fasion-annotation-train.csv'
args.images_dir_test = 'data/fasion-dataset/train'
for n, img_pair in enumerate(os.listdir(in_folder)):
m = re.match(r'([A-Za-z0-9_]*.jpg)_([A-Za-z0-9_]*.jpg)', img_pair)
fr = m.groups()[0]
to = m.groups()[1]
gen_img = imread(os.path.join(in_folder, img_pair))
gen_img = gen_img[:, (2 * args.image_size[1]):]
df = pd.read_csv(args.annotations_file_test, sep=':')
ano_fr = df[df['name'] == fr].iloc[0]
ano_to = df[df['name'] == to].iloc[0]
kp_fr = pose_utils.load_pose_cords_from_strings(ano_fr['keypoints_y'],
ano_fr['keypoints_x'])
kp_to = pose_utils.load_pose_cords_from_strings(ano_to['keypoints_y'],
ano_to['keypoints_x'])
mask = pose_transform.pose_masks(kp_to,
img_size=args.image_size).astype(bool)
mask = np.array(reduce(np.logical_or, list(mask)))
mask = mask.astype('float')
pose_fr, _ = pose_utils.draw_pose_from_cords(kp_fr, args.image_size)
pose_to, _ = pose_utils.draw_pose_from_cords(kp_to, args.image_size)
cur_folder = os.path.join(out_folder, str(n))
if not os.path.exists(cur_folder):
os.makedirs(cur_folder)
