def __init__(self, opt, split='train'):
assert split in {'train', 'val', 'test'}
super(Dataset, self).__init__()
self.opt = opt
self.split = split
if split in {'train', 'val'}:
self.label = io.load_json(opt.fn_label)['label']
self.id_list = io.load_json(opt.fn_split)['train' if split=='train' else 'test']
self.image_dir = opt.img_dir
else:
self.label = None
self.image_dir = opt.test_dir
self.id_list = [fn[0:-4] for fn in os.listdir(self.image_dir)]
if split == 'train':
self.transform = transforms.Compose([
transforms.RandomResizedCrop(opt.crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5,0.5,0.5],std=[0.5,0.5,0.5])
])
else:
self.transform = transforms.Compose([
transforms.Resize(opt.rescale_size),
transforms.CenterCrop(opt.crop_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5,0.5,0.5],std=[0.5,0.5,0.5])
])
if opt.debug:
self.id_list = self.id_list[0:64]
def visualize_seg_map():
num_sample = 1000
output_dir = 'temp/seg_map'
io.mkdir_if_missing(output_dir)
samples = io.load_json(design_root + 'Label/ca_samples.json')
split = io.load_json(design_root + 'Split/ca_gan_split_trainval.json')
id_list = split['train'] + split['test']
id_list = [s_id for s_id in id_list if samples[s_id]['cloth_type'] == 3]
seg_dir_list = [
design_root + 'Img/seg_ca_256/', design_root + 'Img/seg_ca_syn_256/'
]
for i, s_id in enumerate(id_list[0:num_sample]):
s = samples[s_id]
img = image.imread(s['img_path'])
imgs = [img]
for seg_dir in seg_dir_list:
seg = image.imread(seg_dir + s_id + '.bmp') * 20
mask = img * (seg > 0).astype(np.float)
imgs += [seg, mask]
img = image.stitch(imgs, 0)
image.imwrite(img, os.path.join(output_dir, s_id + '.jpg'))
print(i)
def visualize_samples():
num_sample = 10
dir_out = 'temp/attr_example'
io.mkdir_if_missing(dir_out)
samples = io.load_json(design_root + 'Label/ca_samples.json')
attr_label = io.load_data(design_root + 'Label/ca_attr_label.pkl')
attr_entry = io.load_json(design_root + 'Label/attr_entry.json')
id_set = set(samples.keys())
for i, att in enumerate(attr_entry):
print('attribute %d / %d: %s' % (i, len(attr_entry), att['entry']))
dir_att = os.path.join(dir_out, att['entry'])
io.mkdir_if_missing(dir_att)
pos_id_list = [
s_id for s_id, label in attr_label.iteritems() if label[i] == 1
]
np.random.shuffle(pos_id_list)
for s_id in pos_id_list[0:num_sample]:
fn_src = samples[s_id]['img_path']
fn_tar = os.path.join(dir_att, 'pos_' + s_id + '.jpg')
shutil.copyfile(fn_src, fn_tar)
neg_id_list = list(id_set - set(pos_id_list))
np.random.shuffle(neg_id_list)
for s_id in neg_id_list[0:num_sample]:
fn_src = samples[s_id]['img_path']
fn_tar = os.path.join(dir_att, 'neg_' + s_id + '.jpg')
shutil.copyfile(fn_src, fn_tar)
def create_test_pair():
np.random.rand(0)
split = io.load_json(design_root +
'Split/ca_gan_split_trainval_upper.json')
cat_label = io.load_data(design_root + 'Label/ca_cat_label.pkl')
cat_entry = io.load_json(design_root + 'Label/cat_entry.json')
# group samples by category label
cat_to_ids = defaultdict(lambda: [])
for s_id in split['test']:
c = cat_label[s_id]
cat_to_ids[c].append(s_id)
n = 0
pair_list = []
for c, s_list in cat_to_ids.iteritems():
print('[%d/%d] %s: %d samples...' %
(n, len(cat_to_ids), cat_entry[c]['entry'], len(s_list)))
n += 1
s_list_org = [s_id for s_id in s_list]
for i in range(len(s_list) - 1):
j = np.random.randint(i + 1, len(s_list))
temp = s_list[i]
s_list[i] = s_list[j]
s_list[j] = temp
pair_list += zip(s_list_org, s_list)
pair_dict = {s_tar: s_src for s_tar, s_src in pair_list}
io.save_json(pair_dict, design_root + 'Temp/ca_test_tps_pair.json')
io.save_str_list(pair_dict.keys(),
design_root + 'Temp/ca_test_tps_tar.txt')
io.save_str_list(pair_dict.values(),
design_root + 'Temp/ca_test_tps_src.txt')
def create_silhouette():
# DeepFashion
#smpl_pred_dir = 'datasets/DF_Pose/3d/hmr_dfm_v2/pred/'
#output_dir = 'datasets/DF_Pose/Img/silhouette24/'
#image_split = io.load_json('datasets/DF_Pose/Label/image_split_dfm.json')
# Market-1501
smpl_pred_dir = 'datasets/market1501/3d/hmr/pred/'
output_dir = 'datasets/market1501/Images/silhouette24/'
image_split = io.load_json('datasets/market1501/Label/image_split.json')
faces = np.load('scripts/3d/smpl_faces.npy')
vert2part = io.load_json('scripts/3d/smpl_vert_to_bodypart.json')
def _func(face_id):
if face_id == 4294967295:
return 0
else:
verts = faces[face_id]
part_id = vert2part[verts[0]] + 1
return part_id
_vfunc = np.vectorize(_func)
io.mkdir_if_missing(output_dir)
id_list = image_split['train'] + image_split['test']
for sid in tqdm.tqdm(id_list):
pred = io.load_data(smpl_pred_dir + '%s.pkl' % sid)
vis = pred['visibility']
silh = _vfunc(vis).astype(np.uint8)
silh = cv2.medianBlur(silh, 5)
imageio.imwrite(output_dir + '%s.bmp' % sid, silh)
def initialize(self, opt, split):
self.opt = opt
self.root = opt.data_root
self.split = split
print('loading data ...')
samples = io.load_json(os.path.join(opt.data_root, opt.fn_sample))
attr_label = io.load_data(os.path.join(opt.data_root, opt.fn_label))
attr_entry = io.load_json(os.path.join(opt.data_root, opt.fn_entry))
attr_split = io.load_json(os.path.join(opt.data_root, opt.fn_split))
lm_label = io.load_data(os.path.join(opt.data_root, opt.fn_landmark))
self.id_list = attr_split[split]
if opt.max_dataset_size != float('inf'):
self.id_list = self.id_list[0:opt.max_dataset_size]
self.sample_list = [samples[s_id] for s_id in self.id_list]
self.attr_label_list = [attr_label[s_id] for s_id in self.id_list]
self.lm_list = [lm_label[s_id] for s_id in self.id_list]
self.attr_entry = attr_entry
# check data
assert len(self.attr_entry) == len(
self.attr_label_list[0]
) == opt.n_attr, 'Attribute number not match!'
print('dataset created (%d samples)' % len(self))
# get transform
self.to_tensor = transforms.ToTensor()
if opt.image_normalize == 'imagenet':
self.tensor_normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
else:
self.tensor_normalize = transforms.Normalize([0.5, 0.5, 0.5],
[0.5, 0.5, 0.5])
def load_attr_data(self):
opt = self.opt
self.samples = io.load_json(os.path.join(opt.data_root, opt.fn_sample))
self.attr_label = io.load_data(
os.path.join(opt.data_root, opt.fn_label))
self.attr_entry = io.load_json(
os.path.join(opt.data_root, opt.fn_entry))
self.data_loaded = True
def initialize(self, opt, split):
self.opt = opt
self.root = opt.data_root
# get transform
transform_list = []
if opt.resize_or_crop == 'resize':
# only resize image
transform_list.append(transforms.Resize(opt.fine_size, Image.BICUBIC))
elif opt.resize_or_crop == 'resize_and_crop':
# scale and crop
transform_list.append(transforms.Resize(opt.load_size, Image.BICUBIC))
if split == 'train':
transform_list.append(transforms.RandomCrop(opt.fine_size))
transform_list.append(transforms.RandomHorizontalFlip())
else:
transform_list.append(transforms.CenterCrop(opt.fine_size))
transform_list.append(transforms.ToTensor())
if opt.image_normalize == 'imagenet':
transform_list.append(transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]))
else:
transform_list.append(transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]))
self.transform = transforms.Compose(transform_list)
# load sample list
print('loading data ...')
samples = io.load_json(os.path.join(opt.data_root, opt.fn_sample))
attr_label = io.load_data(os.path.join(opt.data_root, opt.fn_label))
attr_entry = io.load_json(os.path.join(opt.data_root, opt.fn_entry))
attr_split = io.load_json(os.path.join(opt.data_root, opt.fn_split))
self.id_list = attr_split[split]
if opt.max_dataset_size != float('inf'):
self.id_list = self.id_list[0:opt.max_dataset_size]
self.sample_list = [samples[s_id] for s_id in self.id_list]
self.attr_label_list = [attr_label[s_id] for s_id in self.id_list]
self.attr_entry = attr_entry
if opt.joint_cat:
cat_label = io.load_data(os.path.join(opt.data_root, opt.fn_cat))
self.cat_list = [cat_label[s_id] for s_id in self.id_list]
if opt.unmatch:
np.random.shuffle(self.sample_list)
# check data
assert len(self.attr_entry) == len(self.attr_label_list[0]) == opt.n_attr, 'Attribute number not match!'
print('dataset created (%d samples)' % len(self))
def create_inner_edge_map():
'''
extract the edges inside the clothing regions
'''
# config
kernel_size = 7
threshold = 0
split = io.load_json(design_root + 'Split/ca_gan_split_trainval.json')
id_list = split['train'] + split['test']
edge_root = design_root + 'Img/edge_ca_256'
seg_root = design_root + 'Img/seg_ca_256'
output_dir = design_root + 'Img/edge_ca_256_inner'
io.mkdir_if_missing(output_dir)
kernel = np.zeros((kernel_size, kernel_size), np.uint8)
k = (kernel_size - 1) / 2
for i in range(kernel_size):
for j in range(kernel_size):
if np.abs(i - k) + np.abs(j - k) <= k:
kernel[i, j] = 1
for i, s_id in enumerate(id_list):
edge = image.imread(os.path.join(edge_root, s_id + '.jpg'),
'grayscale')
seg = image.imread(os.path.join(seg_root, s_id + '.bmp'), 'grayscale')
mask_upper = cv2.erode((seg == 3).astype(np.uint8), kernel)
mask_lower = cv2.erode((seg == 4).astype(np.uint8), kernel)
mask = mask_upper | mask_lower
edge_inner = edge * mask
edge_inner = (edge_inner >= threshold).astype(np.uint8) * edge_inner
image.imwrite(edge_inner, os.path.join(output_dir, s_id + '.jpg'))
print('extracting inner edge %d / %d' % (i, len(id_list)))
# create labels
edge_paths = {
s_id: os.path.join(output_dir, s_id + '.jpg')
for s_id in id_list
}
split_debug = io.load_json(design_root + 'Split/debugca_gan_split.json')
edge_paths_debug = {
s_id: p
for s_id, p in edge_paths.iteritems()
if s_id in split_debug['train'] + split_debug['test']
}
io.save_json(edge_paths, design_root + 'Label/ca_edge_inner_paths.json')
io.save_json(edge_paths_debug,
design_root + 'Label/debugca_edge_inner_paths.json')
def initialize(self, opt, split):
self.opt = opt
self.root = opt.data_root
self.split = split
print('loading data ...')
samples = io.load_json(os.path.join(opt.data_root, opt.fn_sample))
# attr_label = io.load_data(os.path.join(opt.data_root, opt.fn_label))
# attr_entry = io.load_json(os.path.join(opt.data_root, opt.fn_entry))
data_split = io.load_json(os.path.join(opt.data_root, opt.fn_split))
lm_label = io.load_data(os.path.join(opt.data_root, opt.fn_landmark))
seg_paths = io.load_json(os.path.join(opt.data_root, opt.fn_seg_path))
edge_paths = io.load_json(os.path.join(opt.data_root,
opt.fn_edge_path))
# color_paths = io.load_json(os.path.join(opt.data_root, opt.fn_color_path))
flx_seg_paths = io.load_json(
os.path.join(opt.data_root, opt.fn_flx_seg_path))
self.id_list = data_split[split]
# self.attr_entry = attr_entry
if opt.max_dataset_size != float('inf'):
self.id_list = self.id_list[0:opt.max_dataset_size]
self.sample_list = [samples[s_id] for s_id in self.id_list]
# self.attr_label_list = [attr_label[s_id] for s_id in self.id_list]
self.lm_list = [lm_label[s_id] for s_id in self.id_list]
self.seg_path_list = [seg_paths[s_id] for s_id in self.id_list]
self.edge_path_list = [edge_paths[s_id] for s_id in self.id_list]
# self.color_path_list = [color_paths[s_id] for s_id in self.id_list]
self.flx_seg_path_list = [flx_seg_paths[s_id] for s_id in self.id_list]
# check data
# assert len(self.attr_entry) == len(self.attr_label_list[0]) == opt.n_attr, 'Attribute number not match!'
print('dataset created (%d samples)' % len(self))
# get transform
self.to_tensor = transforms.ToTensor()
# use standard normalization, which is different from attribute dataset
# image will be normalized again (under imagenet distribution) before fed into attribute encoder in GAN model
self.tensor_normalize_std = transforms.Normalize([0.5, 0.5, 0.5],
[0.5, 0.5, 0.5])
self.tensor_normalize_imagenet = transforms.Normalize(
[0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
if self.opt.color_jitter:
self.color_jitter = transforms.ColorJitter(brightness=0.3,
contrast=0.3,
saturation=0.3,
hue=0.3)
self.to_pil_image = transforms.ToPILImage()
def create_cloth_edge_map():
'''
create edge map that only contains cloth edge (inside the cloth mask)
'''
# config
mask_dilate = 5
seg_dir = design_root + 'Img/seg_ca_syn_256/'
edge_dir = design_root + 'Img/edge_ca_256/'
output_dir = design_root + 'Img/edge_ca_256_cloth/'
io.mkdir_if_missing(output_dir)
split = io.load_json(design_root +
'Split/ca_gan_split_trainval_upper.json')
id_list = split['train'] + split['test']
for i, s_id in enumerate(id_list):
print('%d/%d' % (i, len(id_list)))
seg_map = image.imread(seg_dir + s_id + '.bmp', 'grayscale')
edge_map = image.imread(edge_dir + s_id + '.jpg', 'grayscale')
assert seg_map.shape == edge_map.shape
mask = ((seg_map == 3) | (seg_map == 4)).astype(np.uint8)
mask = cv2.dilate(mask, kernel=np.ones((mask_dilate, mask_dilate)))
edge_map_cloth = edge_map * mask
image.imwrite(edge_map_cloth, output_dir + s_id + '.jpg')
def create_debug_ca_dataset():
'''
Create a mini subset of Category_and_Attribute data. Assume standard CA index file and label files already exist.
'''
num_train = 10
num_test = 10
same_train_test = True
samples = io.load_json(design_root + 'Label/ca_samples.json')
attr_label = io.load_data(design_root + 'Label/ca_attr_label.pkl')
bbox_label = io.load_data(design_root + 'Label/ca_bbox_label_256.pkl')
lm_label = io.load_data(design_root + 'Label/ca_landmark_label_256.pkl')
if same_train_test:
id_list = samples.keys()[0:num_train]
split = {'train': id_list, 'test': id_list}
else:
id_list = samples.keys()[0:(num_train + num_test)]
split = {'train': id_list[0:num_train], 'test': id_list[num_train::]}
samples = {s_id:samples[s_id] for s_id in id_list}
attr_label = {s_id:attr_label[s_id] for s_id in id_list}
bbox_label = {s_id:bbox_label[s_id] for s_id in id_list}
lm_label = {s_id:lm_label[s_id] for s_id in id_list}
io.save_json(samples, design_root + 'Label/debugca_samples.json')
io.save_data(attr_label, design_root + 'Label/debugca_attr_label.pkl')
io.save_data(bbox_label, design_root + 'Label/debugca_bbox_label.pkl')
io.save_data(lm_label, design_root + 'Label/debugca_landmark_label.pkl')
io.save_json(split, design_root + 'Split/debugca_split.json')
def initialize(self, opt, split):
self.opt = opt
self.root = opt.data_root
if opt.debug:
split = 'debug'
self.split = split
#############################
# load data
#############################
print('loading data ...')
# data split
data_split = io.load_json(os.path.join(opt.data_root, opt.fn_split))
self.pose_label = io.load_data(os.path.join(opt.data_root,
opt.fn_pose))
self.img_dir = os.path.join(opt.data_root, opt.img_dir)
self.seg_dir = os.path.join(opt.data_root, opt.seg_dir)
self.edge_dir = os.path.join(opt.data_root, opt.edge_dir)
#############################
# create index list
#############################
self.id_list = data_split[split]
#############################
# other
#############################
self.tensor_normalize_std = transforms.Normalize([0.5, 0.5, 0.5],
[0.5, 0.5, 0.5])
self.color_jitter = transforms.ColorJitter(brightness=0.3,
contrast=0.3,
saturation=0.3,
hue=0.3)
self.to_pil_image = transforms.ToPILImage()
def create_attribute_label():
img_split = io.load_json('datasets/DF_Pose/Label/image_split_dfm_new.json')
id_list = img_split['train'] + img_split['test']
attr_entry = io.load_str_list(
'datasets/DeepFashion/In-shop/Anno/list_attr_cloth.txt')[2:]
attr_anno = io.load_str_list(
'datasets/DeepFashion/In-shop/Anno/list_attr_items.txt')
attr_anno = attr_anno[2:]
attr_anno = [l.replace('-1', '0').split() for l in attr_anno]
attr_anno = {l[0]: np.array(l[1:], dtype=np.int) for l in attr_anno}
label = {}
for sid in id_list:
s = sid.index('id') + 2
e = s + 8
sid_ori = 'id_' + sid[s:e]
label[sid] = attr_anno[sid_ori]
# remove attribute entries with no positive sample
label_mat = np.array(label.values())
valid_idx = label_mat.sum(axis=0) > 0
print('%d valid attribute entries' % (valid_idx.sum()))
label = {k: v[valid_idx] for k, v in label.iteritems()}
attr_entry = [e for i, e in enumerate(attr_entry) if valid_idx[i]]
attr_label = {'label': label, 'entry': attr_entry}
io.save_data(attr_label, 'datasets/DF_Pose/Label/attr_label.pkl')
def create_attr_label():
# attr_list = io.load_str_list(root + 'Anno/list_attr_cloth.txt')[2::]
# attr_anno = io.load_str_list(root + 'Anno/list_attr_items.txt')[2::]
# fn_out = root + 'Label/attribute_inshop.json'
# num_attr = 463
# n_top = 5
attr_list = io.load_str_list('/data2/ynli/datasets/DeepFashion/Category_and_Attribute/Anno/list_attr_cloth.txt')[2::]
attr_list = [' '.join(a.split()[0:-1]) for a in attr_list]
attr_anno = io.load_str_list('/data2/ynli/datasets/DeepFashion/Category_and_Attribute/Anno/list_attr_img.txt')[2::]
fn_out = root + 'Label/attribute_ca.json'
num_attr = 1000
n_top = 5
# create label data
if not os.path.isfile(fn_out):
attr_data = {}
for line in attr_anno:
line = line.split()
item_id = line[0]
label = [int(c) for c in line[1::]]
assert len(label) == num_attr
attr_data[item_id] = label
io.save_json(attr_data, fn_out)
else:
attr_data = io.load_json(fn_out)
num_sample = len(attr_data)
# most frequent attribute in each attribute type
attr_list_ref = io.load_str_list('/data2/ynli/datasets/DeepFashion/Category_and_Attribute/Anno/list_attr_cloth.txt')[2::]
attr_type = {' '.join(a.split()[0:-1]): a.split()[-1] for a in attr_list_ref}
attr_mat = np.array(attr_data.values(), dtype = float)
attr_count = np.where(attr_mat > 0, 1, 0).sum(axis = 0)
attr_count_type = {}
for i, attr_name in enumerate(attr_list):
t = attr_type[attr_name] if attr_name in attr_type else '-1'
if t not in attr_count_type:
attr_count_type[t] = []
attr_count_type[t].append((attr_name, attr_count[i]))
for t in {'1', '2', '3', '4', '5', '-1'}:
if t not in attr_count_type:
continue
attr_count_list = attr_count_type[t]
attr_count_list.sort(key = lambda x: x[1], reverse = True)
print('attribute type: %s' % t)
for attr_name, count in attr_count_list[0:n_top]:
print('%s: %d (%.1f %%)' % (attr_name, count, 100. * count / num_sample))
print('\n')
def _create_stage_1_net(self, opt):
'''
stage-1 network should be a pretrained pose transfer model.
assume it is a vunet for now
'''
# load options
opt_s1 = argparse.Namespace()
dict_opt_s1 = io.load_json(
os.path.join('checkpoints', opt.which_model_stage_1,
'train_opt.json'))
opt_s1.__dict__.update(dict_opt_s1)
self.opt_s1 = opt_s1
# create model
if opt_s1.which_model_T == 'vunet':
self.netT_s1 = networks.VariationalUnet(
input_nc_dec=self.get_pose_dim(opt_s1.pose_type),
input_nc_enc=self.get_appearance_dim(opt_s1.appearance_type),
output_nc=self.get_output_dim(opt_s1.output_type),
nf=opt_s1.vunet_nf,
max_nf=opt_s1.vunet_max_nf,
input_size=opt_s1.fine_size,
n_latent_scales=opt_s1.vunet_n_latent_scales,
bottleneck_factor=opt_s1.vunet_bottleneck_factor,
box_factor=opt_s1.vunet_box_factor,
n_residual_blocks=2,
norm_layer=networks.get_norm_layer(opt_s1.norm),
activation=nn.ReLU(False),
use_dropout=False,
gpu_ids=opt.gpu_ids,
output_tanh=False,
)
if opt.gpu_ids:
self.netT_s1.cuda()
else:
raise NotImplementedError()
def gather_pose_estimation_result():
'''
We use the state-of-the-art human pose estimation method (https://github.com/tensorboy/pytorch_Realtime_Multi-Person_Pose_Estimation) to get key points
This function is for gathering estimation results.
'''
num_key_p = 18
rst_dir = 'datasets/DeepFashion/Fashion_design/Temp/pose_pkl/'
split = io.load_json(
'datasets/DeepFashion/Fashion_design/Split/ca_gan_split_trainval_upper.json'
)
id_list = split['train'] + split['test']
pose_label = {}
n_fail = 0
for idx, s_id in enumerate(id_list):
print('%d/%d : %s' % (idx, len(id_list), s_id))
fn_pose = rst_dir + s_id + '.pkl'
if not os.path.isfile(fn_pose):
pose_label[s_id] = [[-1, -1] for _ in range(num_key_p)]
n_fail += 1
else:
pose = io.load_data(fn_pose)
assert len(pose) == num_key_p
# p[i][j] = (x, y, score, id) is the j-th keypoints of i-th type.
# we assume that j <= 1, because our image contains single person
pose_label[s_id] = [[p[0][0], p[0][1]] if len(p) > 0 else [-1, -1]
for p in pose]
io.save_data(
pose_label,
'datasets/DeepFashion/Fashion_design/Label/ca_gan_pose_label_256.pkl')
print('%d (out of %d) samples failed' % (n_fail, len(id_list)))
def create_descriptor_seg():
model_id = 'PoseTransfer_7.5'
seg_dir = 'datasets/DF_Pose/Img/seg-lip_df/'
seg_dir_gen = '/data2/ynli/Fashion/fashionHD/checkpoints/%s/test_seg/' % model_id
image_info = io.load_json('temp/patch_matching/label/image_info.json')
id_1 = image_info['id_1']
id_2 = image_info['id_2']
desc_1 = []
desc_2 = []
desc_gen = []
for i in range(num_sample):
fn_1 = seg_dir + id_1[i] + '.bmp'
fn_2 = seg_dir + id_2[i] + '.bmp'
fn_gen = seg_dir_gen + '%s_%s.bmp' % (id_1[i], id_2[i])
desc_1.append(_read_seg(fn_1))
desc_2.append(_read_seg(fn_2))
desc_gen.append(_read_seg(fn_gen))
data_dict_gt = {
'desc_1': np.stack(desc_1),
'desc_2': np.stack(desc_2),
'name': 'gt_seg'
}
data_dict_gen = {
'desc_1': np.stack(desc_1),
'desc_2': np.stack(desc_gen),
'name': 'gen_seg'
}
scipy.io.matlab.savemat('temp/patch_matching/descriptor/desc_gt_seg.mat',
data_dict_gt)
scipy.io.matlab.savemat('temp/patch_matching/descriptor/desc_gen_seg.mat',
data_dict_gen)
def initialize(self, opt, split):
self.opt = opt
self.data_root = opt.data_root
self.split = split
#############################
# set path / load label
#############################
data_split = io.load_json(os.path.join(opt.data_root, opt.fn_split))
self.img_dir = os.path.join(opt.data_root, opt.img_dir)
self.seg_dir = os.path.join(opt.data_root, opt.seg_dir)
self.corr_dir = os.path.join(opt.data_root, opt.corr_dir)
self.pose_label = io.load_data(os.path.join(opt.data_root,
opt.fn_pose))
#############################
# create index list
#############################
self.id_list = data_split[split]
#############################
# other
#############################
if opt.debug:
self.id_list = self.id_list[0:32]
self.tensor_normalize_std = transforms.Normalize([0.5, 0.5, 0.5],
[0.5, 0.5, 0.5])
self.to_pil_image = transforms.ToPILImage()
self.pil_to_tensor = transforms.ToTensor()
self.color_jitter = transforms.ColorJitter(brightness=0.0,
contrast=0.0,
saturation=0.0,
hue=0.2)
def keypoint_guided_tps():
num_sample = 64
pair_list = io.load_json(
'datasets/DF_Pose/Label/pair_split.json')['test'][0:num_sample]
pose_label = io.load_data('datasets/DF_Pose/Label/pose_label.pkl')
image_dir = 'datasets/DF_Pose/Img/img_df/'
seg_dir = 'datasets/DF_Pose/Img/seg-lip_df_revised/'
output_dir = 'temp/patch_matching/output/tps_keypoint/'
io.mkdir_if_missing(output_dir)
tps = cv2.createThinPlateSplineShapeTransformer()
for i, (id_1, id_2) in enumerate(tqdm.tqdm(pair_list)):
kp_1 = np.array(pose_label[id_1][1:14],
dtype=np.float64).reshape(1, -1, 2)
kp_2 = np.array(pose_label[id_2][1:14],
dtype=np.float64).reshape(1, -1, 2)
kp_matches = []
for j in range(kp_1.shape[1]):
if (kp_1[0, j] >= 0).all() and (kp_2[0, j] >= 0).all():
kp_matches.append(cv2.DMatch(j, j, 0))
if len(kp_matches) == 0:
continue
tps.estimateTransformation(kp_2, kp_1, kp_matches)
img_1 = cv2.imread(image_dir + id_1 + '.jpg')
img_2 = cv2.imread(image_dir + id_2 + '.jpg')
img_w = tps.warpImage(img_1)
seg = cv2.imread(seg_dir + id_2 + '.bmp', cv2.IMREAD_GRAYSCALE)
mask = ((seg == 3) | (seg == 7)).astype(img_w.dtype)[:, :, np.newaxis]
img_out = img_w * mask + img_2 * (1 - mask)
cv2.imwrite(output_dir + '%d_%s_%s.jpg' % (i, id_1, id_2), img_out)
cv2.imwrite(output_dir + 'w%d_%s_%s.jpg' % (i, id_1, id_2), img_w)
def create_synthesis_to_CA_index():
'''
create an index map A: img_syn[i] = img_ca[A[i]]
'''
import scipy.io
syn_name_list = io.load_str_list(
'datasets/DeepFashion/Fashion_synthesis/data_release/benchmark/name_list.txt'
)
samples = io.load_json(
'datasets/DeepFashion/Fashion_design/Label/ca_samples.json')
ca_name2idx = {
s['img_path_org'][s['img_path_org'].find('img/')::]: int(s_id[3::])
for s_id, s in samples.iteritems()
}
ca_name2sz = {}
for i, s in enumerate(samples.values()):
img = image.imread(s['img_path_org'])
h, w = img.shape[0:2]
ca_name2sz[s['img_path_org'][s['img_path_org'].find('img/')::]] = (w,
h)
print('load ca image size: %d/%d' % (i, len(samples)))
syn_idx_list = [ca_name2idx[name] for name in syn_name_list]
syn_org_size_list = [ca_name2sz[name] for name in syn_name_list]
data_out = {
'syn2ca_index': syn_idx_list,
'syn2ca_width': [w for w, _ in syn_org_size_list],
'syn2ca_height': [h for _, h in syn_org_size_list]
}
fn_out = 'datasets/DeepFashion/Fashion_synthesis/data_release/benchmark/index_to_Category_and_Attribute.mat'
scipy.io.savemat(fn_out, data_out)
def create_attr_entry():
'''
Create attribute entry list, which contains original 1000 attributes used in Category_and_Attribute benchmark
'''
print('loading data...')
attr_entry_list = io.load_str_list(ca_root +
'Anno/list_attr_cloth.txt')[2::]
attr_label = io.load_data(design_root + 'Label/ca_attr_label.pkl')
split = io.load_json(design_root + 'Split/ca_split.json')
train_ids = set(split['train'])
attr_mat = np.array(
[v for k, v in attr_label.iteritems() if k in train_ids],
dtype=np.float32)
print('computing positive rates')
num_sample = len(train_ids)
pos_rate = attr_mat.sum(axis=0) / num_sample
attr_entry = []
for idx, s in enumerate(attr_entry_list):
s = s.split()
attr_name = ' '.join(s[0:-1])
attr_type = int(s[-1])
attr_entry.append({
'entry': attr_name,
'type': attr_type,
'pos_rate': pos_rate[idx]
})
io.save_json(attr_entry, design_root + 'Label/attr_entry.json')
def create_split():
'''
Create split following the original partition
'''
split_list = io.load_str_list(ca_root + 'Eval/list_eval_partition.txt')[2:]
split = {'train': [], 'val': [], 'test': []}
samples = io.load_json(design_root + 'Label/ca_samples.json')
img2id = {
s['img_path_org'][s['img_path_org'].find('img')::]: s_id
for s_id, s in samples.iteritems()
}
for s in split_list:
img_path, status = s.split()
s_id = img2id[img_path]
split[status].append(s_id)
io.mkdir_if_missing(design_root + 'Split')
io.save_json(split, design_root + 'Split/ca_split.json')
print('create split')
for status in ['train', 'val', 'test']:
print('%s: %d' % (status, len(split[status])))
split_trainval = {
'train': split['train'] + split['val'],
'test': split['test']
}
io.save_json(split_trainval, design_root + 'Split/ca_split_trainval.json')
def create_category_label():
samples = io.load_json(design_root + 'Label/ca_samples.json')
cat_entry_list = io.load_str_list(ca_root +
'Anno/list_category_cloth.txt')[2::]
cat_list = io.load_str_list(ca_root + 'Anno/list_category_img.txt')[2::]
# create category entry
cat_entry = []
for cat_str in cat_entry_list:
cat_name = ' '.join(cat_str.split()[0:-1])
cat_type = int(cat_str.split()[-1])
cat_entry.append({'entry': cat_name, 'type': cat_type})
io.save_json(cat_entry, design_root + 'Label/cat_entry.json')
print('create category entry')
# create category label
img2id = {
s['img_path_org'][s['img_path_org'].find('img')::]: s_id
for s_id, s in samples.iteritems()
}
cat_label = {}
for idx, s in enumerate(cat_list):
s = s.split()
s_id = img2id[s[0]]
cat = int(s[1]) - 1
cat_label[s_id] = cat
io.save_data(cat_label, design_root + 'Label/ca_cat_label.pkl')
print('create category label')
def create_pose_label():
'''
create 18-keypoint pose label. follow the setting in VITON
'''
pose = io.load_data(zalando_root + 'Source/pose.pkl')
split = io.load_json(zalando_root + 'Split/zalando_split.json')
id_list = split['train'] + split['test']
pose_label = {}
for idx, s_id in enumerate(id_list):
print('%d / %d' % (idx, len(id_list)))
subset = pose[s_id + '_0']['subset'] # [i1, i2, ..., in, totial_score, n]
candidate = pose[s_id + '_0']['candidate'] # [[x_i, y_i, score_i, id_i]]
label = []
for i in subset[0][0:-2]:
i = int(i)
if i == -1:
label.append([-1, -1])
else:
x = candidate[i][0]
y = candidate[i][1]
label.append([x, y])
pose_label[s_id] = label
io.save_data(pose_label, zalando_root + 'Label/zalando_pose_label.pkl')
def pad_image_for_segmentation():
'''
resize and padding image for segmentation (using fashionGAN code)
Todo: add inshop version
'''
sz_tar = 256
output_dir = 'datasets/DeepFashion/Fashion_design/Img/img_ca_pad'
io.mkdir_if_missing(output_dir)
samples = io.load_json(design_root + 'Label/ca_samples.json')
split = io.load_json(design_root + 'Split/ca_gan_split_trainval.json')
id_list = split['train'] + split['test']
# update landmark and bbox
lm_label = io.load_data(design_root + 'Label/ca_landmark_label.pkl')
bbox_label = io.load_data(design_root + 'Label/ca_bbox_label.pkl')
lm_label_pad = {}
bbox_label_pad = {}
io.save_str_list(id_list, os.path.join(output_dir, 'img_ca_pad.txt'))
for i, s_id in enumerate(id_list):
img_org = image.imread(samples[s_id]['img_path_org'])
h, w = img_org.shape[0:2]
if h > w:
img = image.resize(img_org, (-1, sz_tar))
scale = 1. * sz_tar / h
else:
img = image.resize(img_org, (sz_tar, -1))
scale = 1. * sz_tar / w
# img = image.pad_square(img, sz_tar, padding_value = 255, mode = 'lefttop')
# image.imwrite(img, os.path.join(output_dir, s_id + '.jpg'))
bbox_label_pad[s_id] = [c * scale for c in bbox_label[s_id]]
lm_label_pad[s_id] = []
for x, y, v in lm_label[s_id]:
lm_label_pad[s_id].append([x * scale, y * scale, v])
print('padding image %d / %d' % (i, len(id_list)))
io.save_data(lm_label_pad,
design_root + 'Label/ca_landmark_label_pad_%d.pkl' % sz_tar)
io.save_data(bbox_label_pad,
design_root + 'Label/ca_bbox_label_pad_%d.pkl' % sz_tar)
def revise_coat_label():
'''
Reivese the segment label of coat(7) and upperbody(3).
'''
import cv2
img_dir = 'datasets/DF_Pose/Img/img_df/'
seg_dir = 'datasets/DF_Pose/Img/seg-lip_df/'
output_dir = 'datasets/DF_Pose/Img/seg-lip_df_revised/'
split = io.load_json('datasets/DF_Pose/Label/split.json')
id_list = split['train'] + split['test']
pid_to_sids = defaultdict(lambda: [])
for sid in id_list:
pid = sid[0:5]
pid_to_sids[pid].append(sid)
print('find %d person ids' % len(pid_to_sids))
n_revised = 0
io.mkdir_if_missing(output_dir)
for i, (pid, sids) in enumerate(pid_to_sids.items()):
seg_0 = cv2.imread(seg_dir + pid + '_1.bmp',
cv2.IMREAD_GRAYSCALE) # try to load frontal image
if (seg_0 is not None) and (7 in seg_0) and (3 in seg_0):
n_revised += 1
img_0 = cv2.imread(img_dir + pid + '_1.jpg')
mask_u = (seg_0 == 3).astype(np.uint8)
mask_c = (seg_0 == 7).astype(np.uint8)
hist_u = cv2.calcHist([img_0], [0, 1, 2], mask_u, [8] * 3,
[0, 256] * 3)
hist_c = cv2.calcHist([img_0], [0, 1, 2], mask_c, [8] * 3,
[0, 256] * 3)
for sid in sids:
if sid == pid + '_1':
shutil.copyfile(seg_dir + sid + '.bmp',
output_dir + sid + '.bmp')
else:
seg_i = cv2.imread(seg_dir + sid + '.bmp',
cv2.IMREAD_GRAYSCALE)
img_i = cv2.imread(img_dir + sid + '.jpg')
mask_u_i = (seg_i == 3).astype(np.uint8)
mask_c_i = (seg_i == 7).astype(np.uint8)
for mask_i in [mask_u_i, mask_c_i]:
if mask_i.any():
hist_i = cv2.calcHist([img_i], [0, 1, 2], mask_i,
[8] * 3, [0, 256] * 3)
if cv2.compareHist(
hist_i, hist_u,
cv2.HISTCMP_CORREL) < cv2.compareHist(
hist_i, hist_c, cv2.HISTCMP_CORREL):
seg_i[mask_i] = 3
else:
seg_i[mask_i] = 7
cv2.imwrite(output_dir + sid + '.bmp', seg_i)
else:
for sid in sids:
shutil.copyfile(seg_dir + sid + '.bmp',
output_dir + sid + '.bmp')
print('%d / %d (%d revised)' % (i, len(pid_to_sids), n_revised))
def create_attribute_label():
'''
Create attribute label using predifined attribute entries
'''
# config
attr_entry = io.load_json(design_root + 'Label/attr_entry.json')
puncs = u'.,!?"%'
trans_table = {ord(c): u' ' for c in puncs}
# load attribute entry
num_attr = len(attr_entry)
item2attr = defaultdict(lambda: [0] * num_attr)
# load language description
desc_list = io.load_json(inshop_root + 'Anno/list_description_inshop.json')
item2desc = {d['item']: d for d in desc_list}
# attribute matching
i_item = 0
for item_id, d in item2desc.iteritems():
color = d['color'].replace('-', ' ')
d_str = ' ' + ' '.join([color] + d['description']) + ' '
d_str = d_str.lower().translate(trans_table)
label = item2attr[item_id]
for idx, att in enumerate(attr_entry):
if ' ' + att['entry'] + ' ' in d_str:
label[idx] = 1
print('extract attribute label: %d / %d' % (i_item, len(item2desc)))
i_item += 1
samples = io.load_json(design_root + 'Label/inshop_samples.json')
attr_label = {
s_id: item2attr[s['item_id']]
for s_id, s in samples.iteritems()
}
io.save_data(attr_label, design_root + 'Label/inshop_attr_label.pkl')
print('create attribute label')
def create_edge_path():
edge_root = design_root + 'Img/edge_ca_256'
samples = io.load_json(design_root + 'Label/ca_samples.json')
split = io.load_json(design_root + 'Split/ca_gan_split_trainval.json')
edge_path = {
s_id: os.path.join(edge_root, s_id + '.jpg')
for s_id in split['train'] + split['test']
}
io.save_json(edge_path, design_root + 'Label/ca_edge_paths.json')
split = io.load_json(design_root + 'Split/debugca_gan_split.json')
edge_path = {
s_id: os.path.join(edge_root, s_id + '.jpg')
for s_id in split['train'] + split['test']
}
io.save_json(edge_path, design_root + 'Label/debugca_edge_paths.json')
def resize_and_pad():
'''
resize the image that its longer side equals to new_size. Then pad the image to have the size [new_size, new_size]
create new pose label at the same time
'''
# config
new_size = 256
img_root = zalando_root + 'Img/img_zalando/'
output_dir = zalando_root + 'Img/img_zalando_%d/' % new_size
split = io.load_json(zalando_root + 'Split/zalando_split.json')
pose_label = io.load_data(zalando_root + 'Label/zalando_pose_label.pkl')
io.mkdir_if_missing(output_dir)
id_list = split['train'] + split['test']
# id_list = id_list[0:10]
new_pose_label = {}
for i, s_id in enumerate(id_list):
print('%d / %d' % (i, len(id_list)))
# resize image
img = cv2.imread(img_root + s_id + '_0.jpg')
w, h = img.shape[1], img.shape[0]
if w < h:
top = 0
bottom = 0
left = (h-w)//2
right = h-w-left
ratio = new_size/h
else:
top = (w-h)//2
bottom = w-h-top
left = 0
right = 0
ratio = new_size/w
img = cv2.copyMakeBorder(img, top, bottom, left, right, borderType=cv2.BORDER_REPLICATE)
img = cv2.resize(img, dsize=(new_size, new_size), interpolation=cv2.INTER_LINEAR)
cv2.imwrite(output_dir + s_id + '_0.jpg', img)
# resize clothing image
img1 = cv2.imread(img_root + s_id + '_1.jpg')
if not (img1.shape[0] == h and img1.shape[1] == w):
img1 = cv2.resize(img1, dsize=(w,h))
img1 = cv2.copyMakeBorder(img1, top, bottom, left, right, borderType=cv2.BORDER_REPLICATE)
img1 = cv2.resize(img1, dsize=(new_size, new_size), interpolation=cv2.INTER_LINEAR)
cv2.imwrite(output_dir + s_id + '_1.jpg', img1)
# modify pose label
pose = pose_label[s_id]
new_pose = [[(p[0]+left)*ratio, (p[1]+top)*ratio] if p != [-1,-1] else [-1,-1] for p in pose]
new_pose_label[s_id] = new_pose
io.save_data(new_pose_label, zalando_root + 'Label/zalando_pose_label_%d.pkl' % new_size)
