def apply_dir(img_path,
save_path,
crop_size=(128, 128),
scale=1,
blur_algos=['clean'],
noise_types=['clean'],
noise_types2=['clean']):
img_list = _get_paths_from_dir(img_dir)
for path in img_list:
rann = ''
env = None
img = util.read_img(env, path)
import uuid
rann = uuid.uuid4().hex
img_crop = random_crop(img, crop_size)
print(img_crop.shape)
#"""
cv2.imwrite(save_path + '/' + rann + 'mask_.png', img_crop * 255)
#"""
img_resize, _ = resize_img(img, crop_size)
print(img_resize.shape)
#"""
cv2.imwrite(save_path + '/' + rann + 'resize_.png', img_resize * 255)
#"""
img_random_resize, _ = random_resize_img(img, crop_size)
print(img_random_resize.shape)
#"""
cv2.imwrite(save_path + '/' + rann + 'random_resize_.png',
img_random_resize * 255)
#"""
img_cutout = cutout(img, img.shape[0] // 2)
print(img_cutout.shape)
#"""
cv2.imwrite(save_path + '/' + rann + 'cutout_.png', img_cutout * 255)
#"""
img_erasing = random_erasing(img)
print(img_erasing.shape)
#"""
cv2.imwrite(save_path + '/' + rann + 'erasing_.png', img_erasing * 255)
#"""
#scale = 4
img_scale, interpol_algo = scale_img(img, scale)
print(img_scale.shape)
#"""
cv2.imwrite(
save_path + '/' + rann + 'scale_' + str(scale) + '_' +
str(interpol_algo) + '_.png', img_scale * 255)
#"""
img_blur, blur_algo, blur_kernel_size = blur_img(img, blur_algos)
print(img_blur.shape)
#"""
cv2.imwrite(
save_path + '/' + rann + 'blur_' + str(blur_kernel_size) + '_' +
str(blur_algo) + '_.png', img_blur * 255)
#cv2.imwrite(save_path+'/blur__.png',img_blur*255)
#"""
img_noise, noise_algo = noise_img(img, noise_types)
#img_noise, noise_algo = noise_img(img_scale, noise_types)
print(img_noise.shape)
#"""
cv2.imwrite(
save_path + '/' + rann + 'noise_' + str(noise_algo) + '_.png',
img_noise * 255)
#"""
img_noise2, noise_algo2 = noise_img(img_noise, noise_types2)
print(img_noise2.shape)
#"""
cv2.imwrite(
save_path + '/' + rann + 'noise2_' + str(noise_algo2) + '_.png',
img_noise2 * 255)
#"""
img_rrot, angle = random_rotate(img)
print(img_rrot.shape)
#"""
cv2.imwrite(save_path + '/' + rann + 'rrot_' + str(angle) + '_.png',
img_rrot * 255)
#"""
print('Finished')
def __getitem__(self, idx):
args = self.args
annot = self.annotations[idx]
t_s = self._getT(annot["subject"]["bbox"], annot["object"]["bbox"])
t_o = self._getT(annot["object"]["bbox"], annot["subject"]["bbox"])
ys0, ys1, xs0, xs1 = annot["subject"]["bbox"]
yo0, yo1, xo0, xo1 = annot["object"]["bbox"]
datum = {
"url": annot["url"],
"_id": annot["_id"],
"subject": {
"name": annot["subject"]["name"],
"embedding": phrase2vec(
annot["subject"]["name"], self.args.max_phrase_len, 300
),
"bbox": np.asarray(
[
ys0 / annot["height"],
ys1 / annot["height"],
xs0 / annot["width"],
xs1 / annot["width"],
],
dtype=np.float32,
),
"t": np.asarray(t_s, dtype=np.float32),
},
"object": {
"name": annot["object"]["name"],
"embedding": phrase2vec(
annot["object"]["name"], self.args.max_phrase_len, 300
),
"bbox": np.asarray(
[
yo0 / annot["height"],
yo1 / annot["height"],
xo0 / annot["width"],
xo1 / annot["width"],
],
dtype=np.float32,
),
"t": np.asarray(t_o, dtype=np.float32),
},
"label": np.asarray([[annot["label"]]], dtype=np.float32),
"predicate": onehot(args.predicate_categories.index(annot["predicate"]), 9),
"predicate_name": annot["predicate"],
}
if self.split == "test":
del datum["label"]
if self.load_image:
img = read_img(annot["url"], self.args.imagepath)
ih, iw = img.shape[:2]
if "train" in self.split:
t_bbox = transforms.Compose(
[
transforms.ToPILImage("RGB"),
transforms.Pad(4, padding_mode="edge"),
transforms.RandomResizedCrop(32, scale=(0.75, 0.85)),
transforms.ToTensor(),
]
)
else:
t_bbox = transforms.Compose(
[
transforms.ToPILImage("RGB"),
transforms.Pad(4, padding_mode="edge"),
transforms.CenterCrop(32),
transforms.ToTensor(),
]
)
bbox_mask = np.stack(
[
self._getDualMask(ih, iw, annot["subject"]["bbox"], 32).astype(
np.uint8
),
self._getDualMask(ih, iw, annot["object"]["bbox"], 32).astype(
np.uint8
),
np.zeros((32, 32), dtype=np.uint8),
],
2,
)
bbox_mask = t_bbox(bbox_mask)[:2].float() / 255.0
datum["bbox_mask"] = bbox_mask
union_bbox = self.enlarge(
self._getUnionBBox(
annot["subject"]["bbox"], annot["object"]["bbox"], ih, iw
),
1.25,
ih,
iw,
)
if "train" in self.split:
t_bboximg = transforms.Compose(
[
transforms.ToPILImage("RGB"),
transforms.RandomResizedCrop(224, scale=(0.75, 0.85)),
transforms.ColorJitter(0.1, 0.1, 0.1, 0.05),
transforms.ToTensor(),
]
)
else:
t_bboximg = transforms.Compose(
[
transforms.ToPILImage("RGB"),
transforms.CenterCrop(224),
transforms.ToTensor(),
]
)
bbox_img = t_bboximg(self._getAppr(img, union_bbox))
datum["bbox_img"] = bbox_img
if "train" in self.split:
t_fullimg = transforms.Compose(
[
transforms.ToPILImage("RGB"),
transforms.ColorJitter(0.1, 0.1, 0.1, 0.05),
transforms.ToTensor(),
]
)
else:
t_fullimg = transforms.Compose(
[transforms.ToPILImage("RGB"), transforms.ToTensor(),]
)
if self.args.model == "vipcnn":
img_size = 400
elif self.args.model == "pprfcn":
img_size = 720
else:
img_size = 224
datum["full_img"] = t_fullimg(self._getAppr(img, [0, ih, 0, iw], img_size))
return datum
def single_image(img_path,
save_path,
crop_size=(128, 128),
scale=1,
blur_algos=['clean'],
noise_types=['clean'],
noise_types2=['clean']):
env = None
img = util.read_img(
env, img_path
) #read image from path, opens with OpenCV, value ranges from 0 to 1
print(img.shape)
img_crop = random_crop(img, crop_size)
print(img_crop.shape)
#"""
cv2.imwrite(save_path + '/crop_.png', img_crop * 255)
#"""
img_resize, _ = resize_img(img, crop_size)
print(img_resize.shape)
#"""
cv2.imwrite(save_path + '/resize_.png', img_resize * 255)
#"""
img_random_resize, _ = random_resize_img(img, crop_size)
print(img_random_resize.shape)
#"""
cv2.imwrite(save_path + '/random_resize_.png', img_random_resize * 255)
#"""
img_cutout = cutout(img, img.shape[0] // 2)
print(img_cutout.shape)
#"""
cv2.imwrite(save_path + '/cutout_.png', img_cutout * 255)
#"""
img_erasing = random_erasing(img)
print(img_erasing.shape)
#"""
cv2.imwrite(save_path + '/erasing_.png', img_erasing * 255)
#"""
#scale = 4
img_scale, interpol_algo = scale_img(img, scale)
print(img_scale.shape)
#"""
cv2.imwrite(
save_path + '/scale_' + str(scale) + '_' + str(interpol_algo) +
'_.png', img_scale * 255)
#"""
img_blur, blur_algo, blur_kernel_size = blur_img(img, blur_algos)
print(img_blur.shape)
#"""
cv2.imwrite(
save_path + '/blur_' + str(blur_kernel_size) + '_' + str(blur_algo) +
'_.png', img_blur * 255)
#"""
img_noise, noise_algo = noise_img(img, noise_types)
#img_noise, noise_algo = noise_img(img_scale, noise_types)
print(img_noise.shape)
#"""
cv2.imwrite(save_path + '/noise_' + str(noise_algo) + '_.png',
img_noise * 255)
#"""
img_noise2, noise_algo2 = noise_img(img_noise, noise_types2)
print(img_noise2.shape)
#"""
cv2.imwrite(save_path + '/noise2_' + str(noise_algo2) + '_.png',
img_noise2 * 255)
#"""
img_rrot, angle = random_rotate(img)
print(img_rrot.shape)
#"""
cv2.imwrite(save_path + '/rrot_' + str(angle) + '_.png', img_rrot * 255)
#"""
print('Finished')
def __getitem__(self, idx):
args = self.args
annot = self.annotations[idx]
t_s = self._getT(annot['subject']['bbox'], annot['object']['bbox'])
t_o = self._getT(annot['object']['bbox'], annot['subject']['bbox'])
xs0, xs1, ys0, ys1 = annot['subject']['bbox']
xo0, xo1, yo0, yo1 = annot['object']['bbox']
datum = {'url': annot['url'],
'_id': annot['_id'],
'subject': {'name': annot['subject']['name'],
'embedding': phrase2vec(annot['subject']['name'],
self.args.max_phrase_len, 300),
'bbox': np.asarray([xs0 / annot['height'], xs1 / annot['height'],
ys0 / annot['width'], ys1 / annot['width']], dtype=np.float32),
't': np.asarray(t_s, dtype=np.float32)},
'object': {'name': annot['object']['name'],
'embedding': phrase2vec(annot['object']['name'],
self.args.max_phrase_len, 300),
'bbox': np.asarray([xo0 / annot['height'], xo1 / annot['height'],
yo0 / annot['width'], yo1 / annot['width']], dtype=np.float32),
't': np.asarray(t_o, dtype=np.float32)},
'label': np.asarray([[annot['label']]], dtype=np.float32),
'predicate': onehot(args.predicate_categories.index(annot['predicate']), 9),
'predicate_name': annot['predicate'],
}
if self.split == 'test':
del datum['label']
if self.load_image:
img = read_img(annot['url'], self.args.imagepath)
ih, iw = img.shape[:2]
if 'train' in self.split:
t_bbox = transforms.Compose([
transforms.ToPILImage('RGB'),
transforms.Pad(4, padding_mode='edge'),
transforms.RandomResizedCrop(32, scale=(0.75, 0.85)),
transforms.ToTensor(),
])
else:
t_bbox = transforms.Compose([
transforms.ToPILImage('RGB'),
transforms.Pad(4, padding_mode='edge'),
transforms.CenterCrop(32),
transforms.ToTensor(),
])
bbox_mask = np.stack([self._getDualMask(ih, iw, annot['subject']['bbox'], 32).astype(np.uint8),
self._getDualMask(ih, iw, annot['object']['bbox'], 32).astype(np.uint8),
np.zeros((32, 32), dtype=np.uint8)], 2)
bbox_mask = t_bbox(bbox_mask)[:2].float() / 255.
datum['bbox_mask'] = bbox_mask
union_bbox = self.enlarge(self._getUnionBBox(annot['subject']['bbox'], annot['object']['bbox'], ih, iw), 1.25, ih, iw)
if 'train' in self.split:
t_bboximg = transforms.Compose([
transforms.ToPILImage('RGB'),
transforms.RandomResizedCrop(224, scale=(0.75, 0.85)),
transforms.ColorJitter(0.1, 0.1, 0.1, 0.05),
transforms.ToTensor(),
])
else:
t_bboximg = transforms.Compose([
transforms.ToPILImage('RGB'),
transforms.CenterCrop(224),
transforms.ToTensor(),
])
bbox_img = t_bboximg(self._getAppr(img, union_bbox))
datum['bbox_img'] = bbox_img
if 'train' in self.split:
t_fullimg = transforms.Compose([
transforms.ToPILImage('RGB'),
transforms.ColorJitter(0.1, 0.1, 0.1, 0.05),
transforms.ToTensor(),
])
else:
t_fullimg = transforms.Compose([
transforms.ToPILImage('RGB'),
transforms.ToTensor(),
])
if self.args.model == 'vipcnn':
img_size = 400
elif self.args.model == 'pprfcn':
img_size = 720
else:
img_size = 224
datum['full_img'] = t_fullimg(self._getAppr(img, [0, ih, 0, iw], img_size))
return datum
