def produce_target(self, holder):
alig_data = []
if self.scale_range is not None:
max_shape = [
random.randint(*self.scale_range),
random.randint(*self.scale_range)
]
max_shape[0] = int(
np.ceil(max_shape[0] / self.divide_size) * self.divide_size)
max_shape[1] = int(
np.ceil(max_shape[1] / self.divide_size) * self.divide_size)
anchor_tools.reset_anchors((max_shape[1], max_shape[0]))
else:
max_shape = self.input_size
# copy images to the upper left part of the image batch object
for [image, boxes_, klass_] in holder:
try:
# construct an image batch object
image, shift_x, shift_y = Fill_img(image,
target_width=max_shape[0],
target_height=max_shape[1])
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array(
[shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
image = image.astype(np.uint8)
image = cv2.resize(image, (max_shape[0], max_shape[1]))
boxes_[:, 0] *= max_shape[0]
boxes_[:, 1] *= max_shape[1]
boxes_[:, 2] *= max_shape[0]
boxes_[:, 3] *= max_shape[1]
except:
logger.warn('there is an err with but it is was handled')
traceback.print_exc()
image = np.zeros(shape=(max_shape[1], max_shape[0], 3),
dtype=np.float32)
boxes_ = np.array([[0, 0, 100, 100]])
klass_ = np.array([0])
if cfg.TRAIN.vis:
for __box in boxes_:
cv2.rectangle(image, (int(__box[0]), int(__box[1])),
(int(__box[2]), int(__box[3])), (255, 0, 0),
4)
all_boxes, all_labels = anchor_tools.produce_target(boxes_, klass_)
alig_data.append([image, all_boxes, all_labels])
return alig_data
def produce_target(self, holder):
alig_data = []
if self.scale_range is not None:
max_shape = [
random.randint(*self.scale_range),
random.randint(*self.scale_range)
]
max_shape[0] = int(
np.ceil(max_shape[0] / self.divide_size) * self.divide_size)
max_shape[1] = int(
np.ceil(max_shape[1] / self.divide_size) * self.divide_size)
anchor_tools.reset_anchors((max_shape[1], max_shape[0]))
else:
max_shape = self.input_size
# copy images to the upper left part of the image batch object
for [image, boxes_, klass_] in holder:
# construct an image batch object
image, shift_x, shift_y = Fill_img(image,
target_width=max_shape[0],
target_height=max_shape[1])
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array(
[shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
image = image.astype(np.uint8)
image = cv2.resize(image, (max_shape[1], max_shape[0]))
boxes_[:, 0] *= max_shape[1]
boxes_[:, 1] *= max_shape[0]
boxes_[:, 2] *= max_shape[1]
boxes_[:, 3] *= max_shape[0]
if cfg.TRAIN.vis:
for __box in boxes_:
cv2.rectangle(image, (int(__box[0]), int(__box[1])),
(int(__box[2]), int(__box[3])), (255, 0, 0),
4)
all_boxes, all_labels = anchor_tools.produce_target(boxes_, klass_)
if cfg.DATA.channel == 1:
image = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
image = np.expand_dims(image, -1)
alig_data.append([image, all_boxes, all_labels])
return alig_data
def _map_func(self, dp, is_training):
fname, annos = dp
image = cv2.imread(fname, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
labels = annos.split(' ')
boxes = []
for label in labels:
bbox = np.array(label.split(','), dtype=np.float)
boxes.append(bbox)
boxes = np.array(boxes, dtype=np.float)
if is_training:
# if random.uniform(0, 1)>0.7:
image, boxes = Random_scale_withbbox(
image,
boxes,
target_shape=[cfg.MODEL.hin, cfg.MODEL.win],
jitter=0.3)
if random.uniform(0, 1) > 0.5:
image, boxes = Random_flip(image, boxes)
if random.uniform(0, 1) > 0.5:
image = self.color_augmentor(image)
if random.uniform(0, 1) > 0.5:
image = Pixel_jitter(image, 15)
if random.uniform(0, 1) > 0.8:
image = Gray_aug(image)
image, shift_x, shift_y = Fill_img(image,
target_width=cfg.MODEL.win,
target_height=cfg.MODEL.hin)
# boxes[:, 0:4] = boxes[:, 0:4] + np.array([shift_x, shift_y, shift_x, shift_y], dtype='float32')
boxes[:, 0] += shift_x
boxes[:, 1] += shift_y
boxes[:, 2] += shift_x
boxes[:, 3] += shift_y
boxes[boxes[:, 4] >= 0, 4] += shift_x
boxes[boxes[:, 5] >= 0, 5] += shift_y
boxes[boxes[:, 6] >= 0, 6] += shift_x
boxes[boxes[:, 7] >= 0, 7] += shift_y
boxes[boxes[:, 8] >= 0, 8] += shift_x
boxes[boxes[:, 9] >= 0, 9] += shift_y
boxes[boxes[:, 10] >= 0, 10] += shift_x
boxes[boxes[:, 11] >= 0, 11] += shift_y
boxes[boxes[:, 12] >= 0, 12] += shift_x
boxes[boxes[:, 13] >= 0, 13] += shift_y
h, w, _ = image.shape
boxes[:, 0] /= w
boxes[:, 1] /= h
boxes[:, 2] /= w
boxes[:, 3] /= h
boxes[boxes[:, 4] >= 0, 4] /= w
boxes[boxes[:, 5] >= 0, 5] /= h
boxes[boxes[:, 6] >= 0, 6] /= w
boxes[boxes[:, 7] >= 0, 7] /= h
boxes[boxes[:, 8] >= 0, 8] /= w
boxes[boxes[:, 9] >= 0, 9] /= h
boxes[boxes[:, 10] >= 0, 10] /= w
boxes[boxes[:, 11] >= 0, 11] /= h
boxes[boxes[:, 12] >= 0, 12] /= w
boxes[boxes[:, 13] >= 0, 13] /= h
image = cv2.resize(image, (cfg.MODEL.win, cfg.MODEL.hin))
image = image.astype(np.uint8)
### cover the small faces with invisible landmarks
boxes_clean = []
for i in range(boxes.shape[0]):
box = boxes[i]
if box[4] < 0:
image[int(box[1] * cfg.MODEL.hin):int(box[3] * cfg.MODEL.hin),
int(box[0] *
cfg.MODEL.hin):int(box[2] * cfg.MODEL.hin
), :] = cfg.DATA.PIXEL_MEAN
else:
boxes_clean.append([
box[1], box[0], box[3], box[2], box[5], box[4], box[7],
box[6], box[9], box[8], box[11], box[10], box[13], box[12]
])
boxes = np.array(boxes_clean)
# tmp = boxes[:, ::2]
# boxes[:, ::2] = boxes[:, 1::2]
# boxes[:, 1::2] = tmp
if cfg.TRAIN.vis:
for i in range(boxes.shape[0]):
box = boxes[i]
color = (255, 0, 0)
thickness = 2
radius = 2
cv2.rectangle(
image,
(int(box[1] * cfg.MODEL.win), int(box[0] * cfg.MODEL.hin)),
(int(box[3] * cfg.MODEL.win), int(box[2] * cfg.MODEL.hin)),
color, thickness)
for point_x, point_y in zip(box[5::2], box[4::2]):
if point_x > 0 and point_y > 0:
cv2.circle(image, (int(point_x * cfg.MODEL.win),
int(point_y * cfg.MODEL.hin)),
radius, color, thickness)
reg_targets, matches = self.produce_target(boxes)
image = image.astype(np.float32)
# if reg_targets.shape[0] > 0:
# reg_targets = reg_targets[:, :4]
return image, reg_targets, matches
def _map_func(self, dp, is_training):
"""Data augmentation function."""
####customed here
try:
fname, annos = dp
image = cv2.imread(fname, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
labels = annos.split(' ')
boxes = []
for label in labels:
bbox = np.array(label.split(','), dtype=np.float)
boxes.append([bbox[0], bbox[1], bbox[2], bbox[3], bbox[4]])
boxes = np.array(boxes, dtype=np.float)
if is_training:
sample_dice = random.uniform(0, 1)
if sample_dice > 0.8 and sample_dice <= 1:
image, boxes = Random_scale_withbbox(
image,
boxes,
target_shape=[cfg.DATA.hin, cfg.DATA.win],
jitter=0.3)
elif sample_dice > 0.4 and sample_dice <= 0.8:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, boxes_, klass_ = dsfd_aug(image, boxes_, klass_)
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
else:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, boxes_, klass_ = baidu_aug(image, boxes_, klass_)
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
if random.uniform(0, 1) > 0.5:
image, boxes = Random_flip(image, boxes)
if random.uniform(0, 1) > 0.5:
image = self.color_augmentor(image)
else:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, shift_x, shift_y = Fill_img(image,
target_width=cfg.DATA.win,
target_height=cfg.DATA.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array(
[shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
image = image.astype(np.uint8)
image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin))
boxes_[:, 0] *= cfg.DATA.win
boxes_[:, 1] *= cfg.DATA.hin
boxes_[:, 2] *= cfg.DATA.win
boxes_[:, 3] *= cfg.DATA.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
if boxes.shape[0] == 0 or np.sum(image) == 0:
boxes_ = np.array([[0, 0, 100, 100]])
klass_ = np.array([0])
else:
boxes_ = np.array(boxes[:, 0:4], dtype=np.float32)
klass_ = np.array(boxes[:, 4], dtype=np.int64)
except:
logger.warn('there is an err with %s' % fname)
traceback.print_exc()
image = np.zeros(shape=(cfg.DATA.hin, cfg.DATA.win, 3),
dtype=np.float32)
boxes_ = np.array([[0, 0, 100, 100]])
klass_ = np.array([0])
return image, boxes_, klass_
def _map_func(self, dp, is_training):
"""Data augmentation function."""
####customed here
try:
fname, annos = dp
image = cv2.imread(fname, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
labels = annos.split(' ')
boxes = []
for label in labels:
bbox = np.array(label.split(','), dtype=np.float)
boxes.append([bbox[0], bbox[1], bbox[2], bbox[3], bbox[4]])
boxes = np.array(boxes, dtype=np.float)
img = image
# if is_training:
#
# height, width = img.shape[0], img.shape[1]
# c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32)
# if 0:
# input_h = (height | self.opt.pad) + 1
# input_w = (width | self.opt.pad) + 1
# s = np.array([input_w, input_h], dtype=np.float32)
# else:
# s = max(img.shape[0], img.shape[1]) * 1.0
# input_h, input_w = cfg.DATA.hin, cfg.DATA.win
#
#
#
# flipped=False
# if 1:
# if 1:
# s = s * np.random.choice(np.arange(0.6, 1.4, 0.1))
# w_border = self._get_border(128, img.shape[1])
# h_border = self._get_border(128, img.shape[0])
# c[0] = np.random.randint(low=w_border, high=img.shape[1] - w_border)
# c[1] = np.random.randint(low=h_border, high=img.shape[0] - h_border)
#
# if np.random.random() < 0.5:
# flipped=True
# img = img[:, ::-1, :]
# c[0] = width - c[0] - 1
#
#
# trans_output = get_affine_transform(c, s, 0, [input_w, input_h])
#
#
# inp = cv2.warpAffine(img, trans_output,
# (input_w, input_h),
# flags=cv2.INTER_LINEAR)
#
# boxes_ = boxes[:,:4]
# klass_ = boxes[:,4:5]
#
#
# boxes_refine=[]
# for k in range(boxes_.shape[0]):
# bbox = boxes_[k]
#
# cls_id = klass_[k]
# if flipped:
# bbox[[0, 2]] = width - bbox[[2, 0]] - 1
# bbox[:2] = affine_transform(bbox[:2], trans_output)
# bbox[2:] = affine_transform(bbox[2:], trans_output)
# bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, input_w - 1)
# bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, input_h - 1)
#
# boxes_refine.append(bbox)
#
# boxes_refine=np.array(boxes_refine)
#
# image = inp.astype(np.uint8)
#
# if random.uniform(0, 1) > 0.5:
# image =self.color_augmentor(image)
# # if random.uniform(0, 1) > 0.5:
# # image =pixel_jitter(image,15)
# image = image.astype(np.uint8)
#
# boxes = np.concatenate([boxes_refine, klass_], axis=1)
if is_training:
if 1:
height, width = img.shape[0], img.shape[1]
c = np.array([img.shape[1] / 2., img.shape[0] / 2.],
dtype=np.float32)
if 0:
input_h = (height | self.opt.pad) + 1
input_w = (width | self.opt.pad) + 1
s = np.array([input_w, input_h], dtype=np.float32)
else:
s = max(img.shape[0], img.shape[1]) * 1.0
input_h, input_w = cfg.DATA.hin, cfg.DATA.win
flipped = False
if 1:
if 1:
s = s * np.random.choice(np.arange(0.6, 1.4, 0.1))
w_border = self._get_border(128, img.shape[1])
h_border = self._get_border(128, img.shape[0])
c[0] = np.random.randint(low=w_border,
high=img.shape[1] -
w_border)
c[1] = np.random.randint(low=h_border,
high=img.shape[0] -
h_border)
if np.random.random() < 0.5:
flipped = True
img = img[:, ::-1, :]
c[0] = width - c[0] - 1
trans_output = get_affine_transform(
c, s, 0, [input_w, input_h])
inp = cv2.warpAffine(img,
trans_output, (input_w, input_h),
flags=cv2.INTER_LINEAR)
boxes_ = boxes[:, :4]
klass_ = boxes[:, 4:5]
boxes_refine = []
for k in range(boxes_.shape[0]):
bbox = boxes_[k]
cls_id = klass_[k]
if flipped:
bbox[[0, 2]] = width - bbox[[2, 0]] - 1
bbox[:2] = affine_transform(bbox[:2], trans_output)
bbox[2:] = affine_transform(bbox[2:], trans_output)
bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, input_w - 1)
bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, input_h - 1)
boxes_refine.append(bbox)
boxes_refine = np.array(boxes_refine)
image = inp.astype(np.uint8)
if random.uniform(0, 1) > 0.5:
image = self.color_augmentor(image)
# if random.uniform(0, 1) > 0.5:
# image =pixel_jitter(image,15)
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_refine, klass_], axis=1)
# else:
#
# ##### different ratios
# sample_dice = random.uniform(0, 1)
# if sample_dice > 0.8 and sample_dice <= 1:
# image, boxes = Random_scale_withbbox(image, boxes, target_shape=[cfg.DATA.hin, cfg.DATA.win],
# jitter=0.5)
#
# elif sample_dice > 0.4 and sample_dice <= 0.8:
# boxes_ = boxes[:, 0:4]
# klass_ = boxes[:, 4:]
#
# image, boxes_, klass_ = dsfd_aug(image, boxes_, klass_)
#
# image = image.astype(np.uint8)
# boxes = np.concatenate([boxes_, klass_], axis=1)
# else:
# boxes_ = boxes[:, 0:4]
# klass_ = boxes[:, 4:]
# image, boxes_, klass_ = baidu_aug(image, boxes_, klass_)
#
# image = image.astype(np.uint8)
# boxes = np.concatenate([boxes_, klass_], axis=1)
#
# if random.uniform(0, 1) > 0.5:
# image, boxes = Random_flip(image, boxes)
#
# # if random.uniform(0, 1) > 0.3:
# # boxes_ = boxes[:, 0:4]
# # klass_ = boxes[:, 4:]
# # angel = random.uniform(-45, 45)
# # image, boxes_ = Rotate_with_box(image, angel, boxes_)
# # boxes = np.concatenate([boxes_, klass_], axis=1)
#
# if random.uniform(0, 1) > 0.5:
# image = self.color_augmentor(image)
# # if random.uniform(0, 1) > 0.5:
# # image = pixel_jitter(image, 15)
#
# boxes_ = boxes
#
# image, shift_x, shift_y = Fill_img(image, target_width=cfg.DATA.win, target_height=cfg.DATA.hin)
# boxes_[:, 0:4] = boxes_[:, 0:4] + np.array([shift_x, shift_y, shift_x, shift_y], dtype='float32')
# h, w, _ = image.shape
# boxes_[:, 0] /= w
# boxes_[:, 1] /= h
# boxes_[:, 2] /= w
# boxes_[:, 3] /= h
# image = image.astype(np.uint8)
# image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin))
#
# boxes_[:, 0] *= cfg.DATA.win
# boxes_[:, 1] *= cfg.DATA.hin
# boxes_[:, 2] *= cfg.DATA.win
# boxes_[:, 3] *= cfg.DATA.hin
# image = image.astype(np.uint8)
# boxes = boxes_
else:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, shift_x, shift_y = Fill_img(image,
target_width=cfg.DATA.win,
target_height=cfg.DATA.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array(
[shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
image = image.astype(np.uint8)
image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin))
boxes_[:, 0] *= cfg.DATA.win
boxes_[:, 1] *= cfg.DATA.hin
boxes_[:, 2] *= cfg.DATA.win
boxes_[:, 3] *= cfg.DATA.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
if boxes.shape[0] == 0 or np.sum(image) == 0:
boxes_ = np.array([[0, 0, -1, -1]])
klass_ = np.array([0])
else:
boxes_ = np.array(boxes[:, 0:4], dtype=np.float32)
klass_ = np.array(boxes[:, 4], dtype=np.int64)
except:
logger.warn('there is an err with %s' % fname)
traceback.print_exc()
image = np.zeros(shape=(cfg.DATA.hin, cfg.DATA.win, 3),
dtype=np.uint8)
boxes_ = np.array([[0, 0, -1, -1]])
klass_ = np.array([0])
return image, boxes_, klass_
def _map_func(self, dp, is_training):
fname, annos = dp
image = cv2.imread(fname, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
labels = annos.split(' ')
boxes = []
for label in labels:
bbox = np.array(label.split(','), dtype=np.float)
boxes.append([bbox[0], bbox[1], bbox[2], bbox[3], bbox[4]])
boxes = np.array(boxes, dtype=np.float)
#########random scale
############## becareful with this func because there is a Infinite loop in its body
if is_training:
# random_index=random.uniform(0, 1)
# if random_index>0.7:
# image, boxes = Random_scale_withbbox(image, boxes, target_shape=[cfg.MODEL.hin, cfg.MODEL.win],
# jitter=0.3)
# elif random_index<0.3 and random_index<=0.7:
# boxes_ = boxes[:, 0:4]
# klass_ = boxes[:, 4:]
# image, boxes_, klass_ = baidu_aug(image, boxes_, klass_)
# image = image.astype(np.uint8)
# boxes = np.concatenate([boxes_, klass_], axis=1)
# else:
# boxes_ = boxes[:, 0:4]
# klass_ = boxes[:, 4:]
# image, boxes_, klass_ = dsfd_aug(image, boxes_, klass_)
# image = image.astype(np.uint8)
# boxes = np.concatenate([boxes_, klass_], axis=1)
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
if random.uniform(0, 1) > 0.5:
image, shift_x, shift_y = Fill_img(image,
target_width=cfg.MODEL.win,
target_height=cfg.MODEL.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array(
[shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
interp_methods = [
cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA,
cv2.INTER_NEAREST, cv2.INTER_LANCZOS4
]
interp_method = random.choice(interp_methods)
image = cv2.resize(image, (cfg.MODEL.win, cfg.MODEL.hin),
interpolation=interp_method)
boxes_[:, 0] *= cfg.MODEL.win
boxes_[:, 1] *= cfg.MODEL.hin
boxes_[:, 2] *= cfg.MODEL.win
boxes_[:, 3] *= cfg.MODEL.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
# if random.uniform(0, 1) > 0.5:
# image, boxes = Random_flip(image, boxes)
if random.uniform(0, 1) > 0.5:
image = self.color_augmentor(image)
if random.uniform(0, 1) > 0.5:
image = Pixel_jitter(image, 15)
if random.uniform(0, 1) > 0.8:
image = Gray_aug(image)
else:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, shift_x, shift_y = Fill_img(image,
target_width=cfg.MODEL.win,
target_height=cfg.MODEL.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array(
[shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
image = cv2.resize(image, (cfg.MODEL.win, cfg.MODEL.hin))
boxes_[:, 0] *= cfg.MODEL.win
boxes_[:, 1] *= cfg.MODEL.hin
boxes_[:, 2] *= cfg.MODEL.win
boxes_[:, 3] *= cfg.MODEL.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
###cove the small faces
boxes_clean = []
for i in range(boxes.shape[0]):
box = boxes[i]
if (box[3] - box[1]) * (box[2] -
box[0]) < cfg.DATA.cover_small_face:
image[int(box[1]):int(box[3]),
int(box[0]):int(box[2]), :] = cfg.DATA.PIXEL_MEAN
else:
boxes_clean.append([
box[1] / cfg.MODEL.hin, box[0] / cfg.MODEL.win,
box[3] / cfg.MODEL.hin, box[2] / cfg.MODEL.win, box[4]
]) # convert xyxy to yxyx
boxes = np.array(boxes_clean)
if cfg.TRAIN.vis:
for box in boxes:
cv2.rectangle(
image,
(int(box[1] * cfg.MODEL.win), int(box[0] * cfg.MODEL.hin)),
(int(box[3] * cfg.MODEL.win), int(box[2] * cfg.MODEL.hin)),
(255, 0, 0), 2)
cv2.putText(
image, str(int(box[4])),
(int(box[1] * cfg.MODEL.win), int(box[0] * cfg.MODEL.hin)),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 0, 0), 2)
reg_targets, cls_targets = self.produce_target(boxes)
image = image.astype(np.float32)
return image, reg_targets, cls_targets
def _map_func(self,dp,is_training):
"""Data augmentation function."""
####customed here
try:
fname, annos = dp
image = cv2.imread(fname, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
labels = annos.split(' ')
boxes = []
for label in labels:
bbox = np.array(label.split(','), dtype=np.float)
boxes.append([bbox[0], bbox[1], bbox[2], bbox[3], bbox[4]])
boxes = np.array(boxes, dtype=np.float)
img=image
if is_training:
###random crop and flip
height, width = img.shape[0], img.shape[1]
c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32)
if 0:
input_h = (height | self.opt.pad) + 1
input_w = (width | self.opt.pad) + 1
s = np.array([input_w, input_h], dtype=np.float32)
else:
s = max(img.shape[0], img.shape[1]) * 1.0
input_h, input_w = cfg.DATA.hin, cfg.DATA.win
flipped = False
if 1:
if 1:
s = s * np.random.choice(np.arange(0.6, 1.4, 0.1))
w_border = self._get_border(128, img.shape[1])
h_border = self._get_border(128, img.shape[0])
c[0] = np.random.randint(low=w_border, high=img.shape[1] - w_border)
c[1] = np.random.randint(low=h_border, high=img.shape[0] - h_border)
if np.random.random() < 0.5:
flipped = True
img = img[:, ::-1, :]
c[0] = width - c[0] - 1
trans_output = get_affine_transform(c, s, 0, [input_w, input_h])
inp = cv2.warpAffine(img, trans_output,
(input_w, input_h),
flags=cv2.INTER_LINEAR)
boxes_ = boxes[:, :4]
klass_ = boxes[:, 4:5]
boxes_refine = []
for k in range(boxes_.shape[0]):
bbox = boxes_[k]
cls_id = klass_[k]
if flipped:
bbox[[0, 2]] = width - bbox[[2, 0]] - 1
bbox[:2] = affine_transform(bbox[:2], trans_output)
bbox[2:] = affine_transform(bbox[2:], trans_output)
bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, input_w - 1)
bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, input_h - 1)
boxes_refine.append(bbox)
boxes_refine = np.array(boxes_refine)
image = inp.astype(np.uint8)
boxes = np.concatenate([boxes_refine, klass_], axis=1)
####random crop and flip
#### pixel level aug
image=self.train_trans(image=image)['image']
####
else:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, shift_x, shift_y = Fill_img(image, target_width=cfg.DATA.win, target_height=cfg.DATA.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array([shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
image = image.astype(np.uint8)
image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin))
boxes_[:, 0] *= cfg.DATA.win
boxes_[:, 1] *= cfg.DATA.hin
boxes_[:, 2] *= cfg.DATA.win
boxes_[:, 3] *= cfg.DATA.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
if boxes.shape[0] == 0 or np.sum(image) == 0:
boxes_ = np.array([[0, 0, -1, -1]])
klass_ = np.array([0])
else:
boxes_ = np.array(boxes[:, 0:4], dtype=np.float32)
klass_ = np.array(boxes[:, 4], dtype=np.int64)
except:
logger.warn('there is an err with %s' % fname)
traceback.print_exc()
image = np.zeros(shape=(cfg.DATA.hin, cfg.DATA.win, 3), dtype=np.uint8)
boxes_ = np.array([[0, 0, -1, -1]])
klass_ = np.array([0])
return image, boxes_, klass_
def _map_func(self,dp,is_training):
"""Data augmentation function."""
####customed here
try:
fname, annos = dp
image = cv2.imread(fname, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
labels = annos.split(' ')
boxes = []
for label in labels:
bbox = np.array(label.split(','), dtype=np.float)
##the augmentor need ymin,xmin,ymax,xmax
boxes.append([bbox[0], bbox[1], bbox[2], bbox[3], bbox[4]])
boxes = np.array(boxes, dtype=np.float)
###clip the bbox for the reason that some bboxs are beyond the image
# h_raw_limit, w_raw_limit, _ = image.shape
# boxes[:, 2] = np.clip(boxes[:, 2], 0, w_raw_limit)
# boxes[:, 3] = np.clip(boxes[:, 3], 0, h_raw_limit)
# boxes[boxes < 0] = 0
#########random scale
############## becareful with this func because there is a Infinite loop in its body
if is_training:
sample_dice = random.uniform(0, 1)
if sample_dice > 0.8 and sample_dice <= 1:
image, boxes = Random_scale_withbbox(image, boxes, target_shape=[cfg.DATA.hin, cfg.DATA.win],
jitter=0.3)
elif sample_dice > 0.4 and sample_dice <= 0.8:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, boxes_, klass_ = dsfd_aug(image, boxes_, klass_)
if random.randint(0,1):
image, shift_x, shift_y = Fill_img(image, target_width=cfg.DATA.win, target_height=cfg.DATA.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array([shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
interp_methods = [cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_NEAREST,
cv2.INTER_LANCZOS4]
interp_method = random.choice(interp_methods)
image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin), interpolation=interp_method)
boxes_[:, 0] *= cfg.DATA.win
boxes_[:, 1] *= cfg.DATA.hin
boxes_[:, 2] *= cfg.DATA.win
boxes_[:, 3] *= cfg.DATA.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
else:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, boxes_, klass_ = baidu_aug(image, boxes_, klass_)
if random.randint(0, 1):
image, shift_x, shift_y = Fill_img(image, target_width=cfg.DATA.win, target_height=cfg.DATA.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array([shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
interp_methods = [cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_NEAREST,
cv2.INTER_LANCZOS4]
interp_method = random.choice(interp_methods)
image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin), interpolation=interp_method)
boxes_[:, 0] *= cfg.DATA.win
boxes_[:, 1] *= cfg.DATA.hin
boxes_[:, 2] *= cfg.DATA.win
boxes_[:, 3] *= cfg.DATA.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
if random.uniform(0, 1) > 0.5:
image, boxes = Random_flip(image, boxes)
# if random.uniform(0, 1) > 0.5:
# image = Pixel_jitter(image, max_=15)
if random.uniform(0, 1) > 0.5:
image = Random_brightness(image, 35)
if random.uniform(0, 1) > 0.5:
image = Random_contrast(image, [0.5, 1.5])
if random.uniform(0, 1) > 0.5:
image = Random_saturation(image, [0.5, 1.5])
# if random.uniform(0, 1) > 0.5:
# a = [3, 5, 7, 9]
# k = random.sample(a, 1)[0]
# image = Blur_aug(image, ksize=(k, k))
if random.uniform(0, 1) > 0.7:
image = Gray_aug(image)
# if random.uniform(0, 1) > 0.7:
# image = Swap_change_aug(image)
# if random.uniform(0, 1) > 0.7:
# boxes_ = boxes[:, 0:4]
# klass_ = boxes[:, 4:]
# angle = random.sample([-90, 90], 1)[0]
# image, boxes_ = Rotate_with_box(image, boxes=boxes_, angle=angle)
# boxes = np.concatenate([boxes_, klass_], axis=1)
else:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, shift_x, shift_y = Fill_img(image, target_width=cfg.DATA.win, target_height=cfg.DATA.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array([shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
image = image.astype(np.uint8)
image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin))
boxes_[:, 0] *= cfg.DATA.win
boxes_[:, 1] *= cfg.DATA.hin
boxes_[:, 2] *= cfg.DATA.win
boxes_[:, 3] *= cfg.DATA.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
if cfg.TRAIN.vis:
for __box in boxes:
cv2.rectangle(image, (int(__box[0]), int(__box[1])),
(int(__box[2]), int(__box[3])), (255, 0, 0), 4)
###cove the small faces
boxes_clean = []
for i in range(boxes.shape[0]):
box = boxes[i]
if (box[3] - box[1]) < cfg.DATA.cover_small_face or (box[2] - box[0]) < cfg.DATA.cover_small_face:
# image[int(box[1]):int(box[3]), int(box[0]):int(box[2]), :] = 0
continue
else:
boxes_clean.append(box)
boxes = np.array(boxes_clean)
if boxes.shape[0] == 0 or np.sum(image) == 0:
boxes_ = np.array([[0, 0, 100, 100]])
klass_ = np.array([0])
else:
boxes_ = np.array(boxes[:, 0:4], dtype=np.float32)
klass_ = np.array(boxes[:, 4], dtype=np.int64)
except:
logger.warn('there is an err with %s' % fname)
traceback.print_exc()
image = np.zeros(shape=(cfg.DATA.hin, cfg.DATA.win, 3), dtype=np.float32)
boxes_ = np.array([[0, 0, 100, 100]])
klass_ = np.array([0])
all_boxes, all_labels = cfg.ANCHOR.achors.produce_target(boxes_, klass_)
return image, all_boxes, all_labels
def _map_func(self, dp, is_training):
"""Data augmentation function."""
####customed here
try:
fname, annos = dp
image = cv2.imread(fname, cv2.IMREAD_COLOR)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
labels = annos.split(' ')
boxes = []
for label in labels:
sp = label.split(',')
if len(sp) > 1:
bbox = np.array(label.split(','), dtype=np.float)
boxes.append([bbox[0], bbox[1], bbox[2], bbox[3], bbox[4]])
boxes = np.array(boxes, dtype=np.float)
if is_training:
annotations = {
'image': image,
'bboxes': boxes,
'category_id': np.ones(len(boxes))
}
augmented = self.aug(**annotations)
image = np.array(augmented['image'])
boxes = np.array(augmented['bboxes'])
h, w, _ = image.shape
# if len(boxes) == 0:
# print('o')
# exit()
# print(np.unique(boxes[:,0] < 0))
# print(np.unique(boxes[:,1] < 0))
#
# boxes[:,1][boxes[:,1] > h] = h-1
# boxes[:,3][boxes[:,3] > w] = w-1
# boxes[:,3] = min(boxes[:,3],h)
# print(np.unique(boxes[:, 1] > w))
# print(np.unique(boxes[:, 3] > h))
# sample_dice = random.uniform(0, 1)
# if sample_dice > 0.8 and sample_dice <= 1:
# image, boxes = Random_scale_withbbox(image, boxes, target_shape=[cfg.DATA.hin, cfg.DATA.win],
# jitter=0.3)
#
# elif sample_dice > 0.4 and sample_dice <= 0.8:
# boxes_ = boxes[:, 0:4]
# klass_ = boxes[:, 4:]
#
# image, boxes_, klass_ = dsfd_aug(image, boxes_, klass_)
#
# image = image.astype(np.uint8)
# boxes = np.concatenate([boxes_, klass_], axis=1)
# else:
# boxes_ = boxes[:, 0:4]
# klass_ = boxes[:, 4:]
# image, boxes_, klass_ = baidu_aug(image, boxes_, klass_)
#
# image = image.astype(np.uint8)
# boxes = np.concatenate([boxes_, klass_], axis=1)
# if random.uniform(0, 1) > 0.5:
# image, boxes = Random_flip(image, boxes)
#
# if random.uniform(0, 1) > 0.5:
# image =self.color_augmentor(image)
else:
boxes_ = boxes[:, 0:4]
klass_ = boxes[:, 4:]
image, shift_x, shift_y = Fill_img(image,
target_width=cfg.DATA.win,
target_height=cfg.DATA.hin)
boxes_[:, 0:4] = boxes_[:, 0:4] + np.array(
[shift_x, shift_y, shift_x, shift_y], dtype='float32')
h, w, _ = image.shape
boxes_[:, 0] /= w
boxes_[:, 1] /= h
boxes_[:, 2] /= w
boxes_[:, 3] /= h
image = image.astype(np.uint8)
image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin))
boxes_[:, 0] *= cfg.DATA.win
boxes_[:, 1] *= cfg.DATA.hin
boxes_[:, 2] *= cfg.DATA.win
boxes_[:, 3] *= cfg.DATA.hin
image = image.astype(np.uint8)
boxes = np.concatenate([boxes_, klass_], axis=1)
if boxes.shape[0] == 0 or np.sum(image) == 0:
boxes_ = np.array([[0, 0, 100, 100]]).astype(np.float32)
klass_ = np.array([0]).astype(np.float32)
else:
boxes_ = np.array(boxes[:, 0:4], dtype=np.float32)
klass_ = np.array(boxes[:, 4], dtype=np.int64)
except:
try:
image, shift_x, shift_y = Fill_img(image,
target_width=cfg.DATA.win,
target_height=cfg.DATA.hin)
image = image.astype(np.uint8)
image = cv2.resize(image, (cfg.DATA.win, cfg.DATA.hin))
image = image.astype(np.uint8)
# image = np.zeros(shape=(cfg.DATA.hin, cfg.DATA.win, 3), dtype=np.float32)
boxes_ = np.array([[0, 0, 100, 100]]).astype(np.float32)
klass_ = np.array([0]).astype(np.float32)
except:
logger.warn('there is an err with %s' % fname)
traceback.print_exc()
return image, boxes_, klass_