def preprocess_data(self, example):
"""
Applies preprocessing step to a single example
"""
sample = tf.io.parse_single_example(example, image_feature_description)
image = tf.image.decode_png(sample["image"])
bbox = tf.cast(tf.io.decode_raw(sample["bbox"], out_type=tf.int64),
dtype=tf.float32)
label = tf.io.decode_raw(sample["label"], out_type=tf.int64)
bbox = to_xyxy(tf.reshape(bbox, (-1, 4)))
if self.dynamic_size:
shape = tf.shape(image)
self.origin_width = shape[1]
self.origin_height = shape[0]
if not self.augment:
image, bbox, label = self.random_crop(image, bbox, label)
image = tf.image.resize(image,
(self.origin_height, self.origin_width))
if self.convert_xywh:
bbox = convert_to_xywh(bbox)
return image, bbox, label
# Data augmentation
image = augmentation.random_adjust_brightness(image)
image = augmentation.random_adjust_contrast(image)
# crop the region contain at least 1 bounding box
has_smallb = has_small_bbox(bbox)
if self.random_cropping and tf.logical_or(has_smallb,
tf.random.uniform(()) > 0.5):
image, bbox, label = self.random_crop(image, bbox, label)
bbox = normalize_bbox(bbox, self.origin_width, self.origin_height)
image, bbox = augmentation.random_flip_horizontal(image, bbox)
if not has_smallb:
image = augmentation.random_gaussian_blur(image, 0.5)
image, image_shape, _ = resize_and_pad_image(image, jitter=None)
w, h = image_shape[0], image_shape[1]
bbox = tf.stack([
bbox[:, 0] * h,
bbox[:, 1] * w,
bbox[:, 2] * h,
bbox[:, 3] * w,
],
axis=-1)
if self.convert_xywh:
bbox = convert_to_xywh(bbox)
return image, bbox, label
def preprocess_data(sample, img_dims=384, pad_flag=True):
"""
Applies preprocessing step to a single sample.
Arguments:
sample: A dict representing a single training sample.
Returns:
image: Resized and padded image with random horizontal flipping applied.
bbox: Bounding boxes with the shape `(num_objects, 4)` where each box is
of the format `[x, y, width, height]`.
class_id: A tensor representing the class id of the objects, having
shape `(num_objects,)`.
"""
jitter = [sample["l_jitter"], sample["u_jitter"]]
image = _parse_image(sample["image"])
if not pad_flag:
image = tf.image.resize(image, [img_dims, img_dims])
bbox = tf.cast(sample["objects"]["bbox"], tf.float32)
class_id = tf.cast(sample["objects"]["label"], dtype=tf.int32)
image, bbox = random_flip_horizontal(image, bbox)
if pad_flag:
image, img_shp, ratio = \
resize_and_pad_image(
image, min_side=sample["min_side"],
max_side=sample["max_side"], jitter=jitter)
else:
image = image / 127.5 - 1.0
img_shp = tf.cast([img_dims, img_dims], tf.float32)
bbox = swap_xy(bbox)
bbox = convert_to_xywh(bbox)
bbox = bbox.numpy()
return image, tf.constant(bbox), class_id, img_shp
def resize_image(sample):
image = _parse_image(sample["image"])
image = tf.image.resize(image, [sample["min_side"], sample["min_side"]])
image = image / 127.5 - 1.0
bbox = tf.cast(swap_xy(sample["objects"]["bbox"]), tf.float32)
bbox = convert_to_xywh(bbox)
class_id = tf.cast(sample["objects"]["label"], dtype=tf.int32)
return image, bbox, class_id
def prepare_data(datasets, annotations, threthoud_pos, threthoud_neg,
save_path):
global COUNT_FACE, COUNT_BACKGROUND
for i in range(len(datasets)):
image_dir, num_of_faces, gts = datasets[i]
gts = convert_to_xywh(ellipse_to_rectangle(num_of_faces, gts))
for gt in gts:
img = crop_image(image_dir, gt)
if len(img) == 0:
continue
a, b, c = img.shape
if a == 0 or b == 0 or c == 0:
continue
COUNT_FACE += 1
path = ''.join(
[save_path, '1/',
str(i), '_',
str(COUNT_FACE), '.jpg'])
cv2.imwrite(path, img)
for candidate in generate_selective_search(image_dir):
x, y, w, h = candidate
ious = []
img = crop_image(image_dir, candidate)
if len(img) == 0:
continue
for gt in gts:
ious.append(
IOU_calculator(x + w / 2, y + h / 2, w, h,
gt[0] + gt[2] / 2, gt[1] + gt[3] / 2, gt[2],
gt[3]))
if max(ious) > threthoud_pos:
COUNT_FACE += 1
path = ''.join(
[save_path, '1/',
str(i), '_',
str(COUNT_FACE), '.jpg'])
cv2.imwrite(path, img)
elif max(ious) < threthoud_neg:
COUNT_BACKGROUND += 1
path = ''.join([
save_path, '0/',
str(i), '_',
str(COUNT_BACKGROUND), '.jpg'
])
cv2.imwrite(path, img)
print(
f"====>>> {i}/{len(datasets)}: Face: {COUNT_FACE}, Background: {COUNT_BACKGROUND}"
)
def _encode_sample(self, gt_boxes, cls_ids):
"""Создает боксы и классифициет таргеты для одиночного сэмпла"""
anchor_boxes = self._anchor_box
gt_boxes = tf.cast(gt_boxes, dtype=tf.float32)
gt_boxes = tf.reshape(gt_boxes, ((1, ) + gt_boxes.shape))
cls_ids = tf.cast(cls_ids, dtype=tf.float32)
matched_gt_idx, positive_mask, ignore_mask = self._match_anchor_boxes(
anchor_boxes, gt_boxes)
gt_boxes = utils.convert_to_xywh(gt_boxes)
box_target = self._compute_box_target(anchor_boxes, gt_boxes)
cls_gt = tf.ones((self.num_boxes, 1), dtype=tf.float32)
cls_bg = tf.cast(tf.equal(cls_gt, 0.), dtype=tf.float32)
label = tf.concat([box_target, cls_gt], axis=1)
label = tf.concat([label, cls_bg], axis=1)
return label
def decode_sample(self, example):
sample = tf.io.parse_single_example(example, image_feature_description)
image = tf.image.decode_png(sample["image"])
bbox = tf.cast(
tf.io.decode_raw(sample["bbox"], out_type=tf.int64), dtype=tf.float32
)
label = tf.io.decode_raw(sample["label"], out_type=tf.int64)
bbox = tf.reshape(bbox, (-1, 4))
shape = tf.shape(image)
self.set_height(shape[0])
self.set_width(shape[1])
shape = tf.cast(shape, tf.float32)
width = shape[1]
height = shape[0]
bbox = tf.stack([
tf.maximum(bbox[:, 0], 0),
tf.maximum(bbox[:, 1], 0),
tf.minimum(bbox[:, 2], width),
tf.minimum(bbox[:, 3], height),
], axis=-1)
if has_small_bbox(width, height, bbox) and tf.random.uniform(()) > 0.5:
image, bbox, label = self.random_crop(image, bbox, label)
if self.augment:
image = random_adjust_brightness(image)
image = random_adjust_contrast(image)
if tf.random.uniform(()) >= 0.8:
image = tf.image.random_hue(image, 0.1)
if tf.random.uniform(()) >= 0.8:
image = tf.image.random_saturation(image, 0.1, 0.5)
bbox = normalize_bbox(bbox, width, height)
image, bbox = random_flip_horizontal(image, bbox, 0.5)
image, image_shape, _ = resize_and_pad_image(image,
self.resize,
self.resize, jitter=None)
w, h = image_shape[0], image_shape[1]
bbox = tf.stack([
bbox[:, 0] * h,
bbox[:, 1] * w,
bbox[:, 2] * h,
bbox[:, 3] * w,
], axis=-1)
if self.iterator and not self.is_iter and tf.random.uniform(()) > 0.5:
image_, bbox_, label_ = self.iterator.get_next()
shape = tf.shape(image)
shape_ = tf.shape(image_)
# mixup
if shape_[0] == shape[0] and shape[1] == shape_[1]:
image = tf.cast(image, tf.float32)
if tf.size(label_) > 0:
bbox = tf.concat([bbox, bbox_], axis=0)
label = tf.concat([label, label_], axis=0)
r = tf.random.uniform((), 0.35, 0.65)
image = image * r + image_ * (1 - r)
# copy-paste
# image, bbox, label = self.copy_paste(
# image, bbox, label,
# image_, bbox_, label_
# )
if self.convert and not self.is_iter:
bbox = convert_to_xywh(bbox)
return image, bbox, label