def __getitem__(self, index):
im_path = self.image_list[index]
im = cv2.cvtColor(cv2.imread(im_path, cv2.IMREAD_COLOR),
cv2.COLOR_BGR2RGB)
roidb = self._load_annotation(self.image_list[index])
gt_inds = np.where(roidb['gt_classes'] != 0)[0]
bboxes = roidb['boxes'][gt_inds, :]
classes = roidb['gt_classes'][gt_inds]
gt_boxes = np.zeros((len(gt_inds), 6), dtype=np.float32)
if self.augment:
transform = Augment(
[
HorizontalFlip(0.5),
VerticalFlip(0.5),
Affine(degree=0, translate=0.1, scale=0., p=0.5),
],
box_mode='xyxyxyxy',
)
im, bboxes = transform(im, bboxes)
mask = mask_valid_boxes(quad_2_rbox(bboxes, 'xywha'), return_mask=True)
bboxes = bboxes[mask]
gt_boxes = gt_boxes[mask]
classes = classes[mask]
for i, bbox in enumerate(bboxes):
gt_boxes[i, :5] = quad_2_rbox(np.array(bbox),
mode='xyxya') # 四点转xyxya（a为角度制）
gt_boxes[i, 5] = classes[i]
## test augmentation
# print(im.shape)
# plot_gt(im, gt_boxes[:,:5], im_path, mode = 'xyxya')
return {'image': im, 'boxes': gt_boxes, 'path': im_path}
def _load_annotation(self, index):
root_dir, img_name = os.path.split(index)
filename = os.path.join(root_dir, img_name[:-5] + '.xml')
boxes, gt_classes = [], []
with open(filename, 'r', encoding='utf-8-sig') as f:
content = f.read()
assert '<objects>' in content, 'Background picture occurred in %s' % filename
objects = content.split('<object>')
info = objects.pop(0)
for obj in objects:
assert len(obj) != 0, 'No onject found in %s' % filename
points = obj[obj.find('<points>') +
8:obj.find('</points>')].split('<point>')[1:]
coors = [x.split('<')[0].split(',') for x in points]
(x1, y1), (x2, y2), (x3, y3), (x4, y4), *_ = coors
x1, y1, x2, y2, x3, y3, x4, y4 = [
x for x in map(eval, [x1, y1, x2, y2, x3, y3, x4, y4])
]
quad_box = np.array([(x1, y1), (x2, y2), (x3, y3), (x4, y4)])
box = quad_2_rbox(quad_box)
boxes.append(box)
label = self.class_mapping(cls_id, self.level)
gt_classes.append(label)
return {'boxes': np.array(boxes), 'gt_classes': np.array(gt_classes)}
def __getitem__(self, index):
# index是索引号,每次进8线程同时进行处理（num_worker），self.image_list[index]从trainval.txt的所有img中提取出遍历当前的图片名
im_path = self.image_list[index] # 获取当前张单张图像的绝对路径
im = cv2.cvtColor(cv2.imread(im_path, cv2.IMREAD_COLOR),
cv2.COLOR_BGR2RGB)
# 由当前的文件提取出xml的信息,dict格式内含两对kv：'boxes'-shape(nt,8)和'gt_classes'-shape(1,nt)
roidb = self._load_annotation(self.image_list[index])
gt_inds = np.where(roidb['gt_classes'] != 0)[0]
nt = len(roidb['boxes'])
# gt_boxes的格式为：xyxyac
gt_boxes = np.zeros((len(gt_inds), 6), dtype=np.float32)
if nt:
bboxes = roidb['boxes'][gt_inds, :]
classes = roidb['gt_classes'][gt_inds]
if self.augment:
transform = Augment(
[
HorizontalFlip(0.5),
VerticalFlip(0.5),
Affine(degree=0, translate=0.1, scale=0.1, p=0.5),
# Grayscale(0.3,p=0.5),
# # Contrast(0.15, p=0.3),
# Sharpen(0.15, p=0.2),
Noise(0.1, p=0.5),
# Gamma(0.2, p=0.4),
# Blur(1.3, p=0.5),
],
box_mode='xywha',
)
im, bboxes = transform(im, bboxes)
gt_boxes[:, :-1] = bboxes
mask = mask_valid_boxes(quad_2_rbox(bboxes, 'xywha'),
return_mask=True)
bboxes = bboxes[mask]
gt_boxes = gt_boxes[mask]
classes = classes[mask]
for i, bbox in enumerate(bboxes):
gt_boxes[i, 5] = classes[i]
gt_boxes = constraint_theta(gt_boxes)
cx, cy, w, h = [gt_boxes[:, x] for x in range(4)]
x1 = cx - 0.5 * w
x2 = cx + 0.5 * w
y1 = cy - 0.5 * h
y2 = cy + 0.5 * h
gt_boxes[:, 0] = x1
gt_boxes[:, 1] = y1
gt_boxes[:, 2] = x2
gt_boxes[:, 3] = y2
## test augmentation
print(im.shape)
plot_gt(im, gt_boxes[:, :5], im_path, mode='xyxya')
return {'image': im, 'boxes': gt_boxes, 'path': im_path}
def __call__(self, img, labels, mode=None):
if random.random() < self.p:
if mode == 'xywha':
labels = rbox_2_quad(labels, mode = 'xywha')
img, labels = random_affine(img, labels,
degree=self.degree,translate=self.translate,
scale=self.scale,shear=self.shear )
labels = quad_2_rbox(labels, mode = 'xywha')
else:
img, labels = random_affine(img, labels,
degree=self.degree,translate=self.translate,
scale=self.scale,shear=self.shear )
return img, labels
def __getitem__(self, index):
im_path = self._image_path_from_index(self.image_list[index])
im = cv2.cvtColor(cv2.imread(im_path, cv2.IMREAD_COLOR), cv2.COLOR_BGR2RGB)
roidb = self._load_pascal_annotation(self.image_list[index])
gt_inds = np.where(roidb['gt_classes'] != 0)[0]
bboxes = roidb['boxes'][gt_inds, :]
classes = roidb['gt_classes'][gt_inds]
if self.random_flip and np.random.rand() >= 0.5:
im = cv2.flip(im, 1, None)
oldxs = bboxes[:, 0::2].copy()
bboxes[:, 0::2] = im.shape[1] - oldxs - 1
gt_boxes = np.empty((len(gt_inds), 6), dtype=np.float32)
for i, bbox in enumerate(bboxes):
gt_boxes[i, :5] = quad_2_rbox(np.array(bbox))
gt_boxes[:, 5] = classes[i]
return {'image': im, 'boxes': gt_boxes}
def __getitem__(self, index):
im_path = self._image_path_from_index(self.image_list[index])
im = cv2.cvtColor(cv2.imread(im_path, cv2.IMREAD_COLOR),
cv2.COLOR_BGR2RGB)
if im is not None:
h, w = im.shape[:2]
gt_path = self._annotation_path_from_index(self.annotation_list[index])
if self.dformat == 'txt':
roidb = self._load_annotation_txt(gt_path, h, w)
elif self.dformat == 'json':
roidb = self._load_annotation_json(gt_path)
else:
raise Exception('Data format not supported!')
gt_inds = np.where(roidb['gt_classes'] != 0)[0]
bboxes = roidb['boxes'][gt_inds, :]
classes = roidb['gt_classes'][gt_inds]
gt_boxes = np.empty((len(gt_inds), 6), dtype=np.float32)
for i, bbox in enumerate(bboxes):
gt_boxes[i, :5] = quad_2_rbox(np.array(bbox))
gt_boxes[:, 5] = classes[i]
return {'image': im, 'boxes': gt_boxes}
def random_affine(img,
targets=(),
degree=10,
translate=.1,
scale=.1,
shear=10):
# torchvision.transforms.RandomAffine(degree=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
# https://medium.com/uruvideo/dataset-augmentation-with-random-homographies-a8f4b44830d4
if targets is None:
targets = []
border = 0 # width of added border (optional)
height = img.shape[0] + border * 2
width = img.shape[1] + border * 2
# Rotation and Scale
R = np.eye(3)
a = random.uniform(-degree, degree)
# # # a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations
s = random.uniform(1 - scale, 1 + scale)
R[:2] = cv2.getRotationMatrix2D(angle=a,
center=(img.shape[1] / 2,
img.shape[0] / 2),
scale=s)
# Translation
T = np.eye(3)
T[0,
2] = random.uniform(-translate, translate
) * img.shape[0] + border # x translation (pixels)
T[1,
2] = random.uniform(-translate, translate
) * img.shape[1] + border # y translation (pixels)
M = T @ R # Combined rotation matrix. ORDER IS IMPORTANT HERE!!
imw = cv2.warpAffine(img,
M[:2],
dsize=(width, height),
flags=cv2.INTER_AREA,
borderValue=(128, 128, 128)) # BGR order borderValue
# Return warped points also
targets[:, [0, 2, 4, 6]] = targets[:, [0, 2, 4, 6]] * M[
0, 0] + targets[:, [1, 3, 5, 7]] * M[0, 1] + M[0, 2]
targets[:, [1, 3, 5, 7]] = targets[:, [0, 2, 4, 6]] * M[
1, 0] + targets[:, [1, 3, 5, 7]] * M[1, 1] + M[1, 2]
for x in range(0, 8, 2):
targets[:, x] = targets[:, x].clip(0, width)
for y in range(1, 8, 2):
targets[:, y] = targets[:, y].clip(0, height)
rboxes = quad_2_rbox(targets)
w = rboxes[:, 2]
h = rboxes[:, 3]
ar = np.maximum(w / (h + 1e-16), h / (w + 1e-16))
mask = (w > 4) & \
(h > 4) & \
(ar < 15) & \
np.array([(x>0).all() for x in targets]) & \
np.array([(x[0::2]<width).all() for x in targets]) & \
np.array([(x[1::2]<height).all() for x in targets])
targets = targets[mask]
return imw, targets, mask
