def __parse_data(self, img_path, anno):
img = cv2.imread(img_path) # H*W*C and C=BGR
img_h, img_w = img.shape[0], img.shape[1]
assert img is not None, 'File Not Found ' + img_path
# permutation = [4, 0, 1, 2, 3]
# idx = np.empty_like(permutation)
# idx[permutation] = np.arange(len(permutation))
# anno = anno[:, idx]
if len(anno) == 0:
img, bboxes = dataAug.Resize((self.img_size, self.img_size),
True)(np.copy(img), np.copy(anno))
return img, bboxes
bboxes = anno[:, [
1, 2, 3, 4, 0
]] # return a copy so it will not change the original anno which we will use inthe next epoch
bboxes[:, :4:2] = bboxes[:, :4:2] * img_w
bboxes[:, 1:4:2] = bboxes[:, 1:4:2] * img_h
bboxes[:, 0:4] = dataAug.xywh2xyxy_np(bboxes[:, 0:4])
bboxes = np.round(bboxes)
if self.augmentation:
img, bboxes = dataAug.RandomHorizontalFilp()(np.copy(img),
np.copy(bboxes))
img, bboxes = dataAug.RandomCrop()(np.copy(img), np.copy(bboxes))
img, bboxes = dataAug.RandomAffine()(np.copy(img), np.copy(bboxes))
bboxes = np.where(bboxes < 0.0, 0.0, bboxes)
bboxes[:, 2] = np.where(bboxes[:, 2] > img_w, img_w, bboxes[:, 2])
bboxes[:, 3] = np.where(bboxes[:, 3] > img_h, img_h, bboxes[:, 3])
sample = self.albu_aug(image=np.copy(img), bboxes=np.copy(bboxes))
img, bboxes = sample['image'], sample['bboxes']
bboxes = np.array(bboxes)
if len(bboxes) == 0:
bboxes = np.array(bboxes)
bboxes = bboxes.reshape(
-1, 5
) # so that it will not error when go to Resize when it is blank
# Apply random erasor
# if len(bboxes) != 0:
# img = self.random_erasor_I(np.copy(img))
# img = self.random_erasor_O(np.copy(img), bboxes)
# Resize the image to target size and transforms it into a color channel(BGR->RGB),
# as well as pixel value normalization([0,1])
img, bboxes = dataAug.Resize((self.img_size, self.img_size),
True)(np.copy(img), np.copy(bboxes))
return img, bboxes
def __parse_annotation(self, annotation):
"""
Data augument.
:param annotation: Image' path and bboxes' coordinates, categories.
ex. [image_path xmin,ymin,xmax,ymax,class_ind xmin,ymin,xmax,ymax,class_ind ...]
:return: Return the enhanced image and bboxes. bbox'shape is [xmin, ymin, xmax, ymax, class_ind]
"""
anno = annotation.strip().split(" ")
img_path = anno[0]
img = cv2.imread(img_path) # H*W*C and C=BGR
assert img is not None, "File Not Found " + img_path
bboxes = np.array(
[list(map(float, box.split(","))) for box in anno[1:]]
)
img, bboxes = dataAug.RandomHorizontalFilp()(
np.copy(img), np.copy(bboxes), img_path
)
img, bboxes = dataAug.RandomCrop()(np.copy(img), np.copy(bboxes))
img, bboxes = dataAug.RandomAffine()(np.copy(img), np.copy(bboxes))
img, bboxes = dataAug.Resize((self.img_size, self.img_size), True)(
np.copy(img), np.copy(bboxes)
)
return img, bboxes
def __data_aug(self, img, bboxes):
img, bboxes = dataAug.RandomHorizontalFilp()(np.copy(img),
np.copy(bboxes))
img, bboxes = dataAug.RandomCrop()(np.copy(img), np.copy(bboxes))
img, bboxes = dataAug.RandomAffine()(np.copy(img), np.copy(bboxes))
img, bboxes = dataAug.Resize((self.img_size, self.img_size),
True)(np.copy(img), np.copy(bboxes))
return img, bboxes