def __call__(self, example):
image = example['image']
im_w, im_h = image.size
if im_w < self.w or im_h < self.h:
if im_h < im_w:
scale = self.h / im_h
image = image.resize((int(0.5 + im_w * scale), self.h))
else:
scale = self.w / im_w
image = image.resize((self.w, int(0.5 + im_h * scale)))
new_boxes = []
for box in example['boxes']:
new_boxes.append(
BoundingBox(box.left * scale, box.top * scale,
box.right * scale, box.bottom * scale,
image.size[0], image.size[1], box.label))
boxes = new_boxes
return {
'image': image,
'boxes': boxes,
'labels': example['labels']
}
return example
def update_task(path, tasks, label_map):
t = pd.read_csv(path, sep='\t', dtype=str)
if t.values.shape[0] == 0:
return
dataset_name = path.split('/')[5]
start = path.find('annotations') + len('annotations') + 1
# print(path, start, path[start:])
task = os.path.join(dataset_name, os.path.splitext(path[start:])[0])
# print(task)
boxes = []
labels = []
ext = t.ext[0] if 'ext' in t else 'jpg'
impath = os.path.join(DIR, dataset_name, 'images', path[start:].replace('.tsv', '.' + ext))
# img = cv2.imread(impath)
# height, width = img.shape[:2]
for _, row in t.iterrows():
if ((float(row.xbr) - float(row.xtl)) >=cfg.width) or ((float(row.ybr) - float(row.ytl))) >= cfg.height:
continue
class_ = row['class']
if class_ not in label_map:
label = 97
else:
label = label_map[class_]
box = BoundingBox(
max(0, float(row.xtl)),
max(0, float(row.ytl)),
min(10000, float(row.xbr)),
min(10000, float(row.ybr)),
-1,
-1,
label)
boxes.append(box)
labels.append(label)
if len(boxes) == 0:
return
impath = os.path.join(DIR, dataset_name, 'images', path[start:].replace('.tsv', '.' + ext))
if not os.path.exists(impath):
print(impath)
1/0
tasks[task] = {
'impath': impath,
'boxes': boxes,
'labels': [int(l) for l in labels],
}
def build_annotations(self):
box_dict = {}
annos = json.loads(open(self.anno_file).read())
for id in annos['imgs']:
boxes = []
for sign_dict in annos['imgs'][id]['objects']:
label = sign_dict['category']
left = int(sign_dict['bbox']['xmin'])
right = int(sign_dict['bbox']['xmax'])
top = int(sign_dict['bbox']['ymin'])
bottom = int(sign_dict['bbox']['ymax'])
box = BoundingBox(left, top, right, bottom, cfg.width,
cfg.height, self.labels.index(label))
boxes.append(box)
if len(boxes) > 0:
box_dict[annos['imgs'][id]['path'].split('/')[-1]] = boxes
return box_dict
def __getitem__(self, idx):
task = self.task_dict[self.tasks[idx]]
image = Image.fromarray(cv2.imread(task['impath'])[...,::-1])
boxes = task['boxes']
labels = task['labels']
width, height = image.size
new_boxes = []
for box in boxes:
box = BoundingBox(
max(0, float(box.left)),
max(0, float(box.top)),
min(width, float(box.right)),
min(height, float(box.bottom)), -1, -1, 0)
new_boxes.append(box)
boxes = new_boxes
example = {
'image': image,
'boxes': boxes,
'labels': labels
}
if not self.test:
example = self.pil_transform(example)
if self.transform is not None:
example = self.transform(example)
if True:
pass
return example
def __call__(self, example):
image = example['image']
boxes = example['boxes']
labels = example['labels']
im_w, im_h = image.size
try:
# print(im_w, im_h)
box = np.random.choice(boxes)
from copy import deepcopy
orig = deepcopy(box)
if (box.right - box.left) > self.w:
x_from = max(0, box.left - self.s)
x_to = x_from + 1
else:
x_from = max(0, box.right - self.w)
x_to = min(im_w - self.w, box.left) + 1
# print(x_from, x_to)
x = np.random.randint(int(x_from), int(x_to))
if (box.bottom - box.top) > self.h:
y_from = max(0, box.top - self.s)
y_to = y_from + 1
else:
y_from = max(0, box.bottom - self.h)
y_to = min(im_h - self.h, box.top) + 1
# print(y_from, y_to)
y = np.random.randint(int(y_from), int(y_to))
except:
print('bbox:', box.left, box.top, box.right, box.bottom)
print('im_h: {}, im_w: {}'.format(im_h, im_w))
print('x:')
print('x_from; max(0,', box.right - self.w)
print('x_to: min', im_w - self.w, box.left)
print(x_from, x_to)
print('y:')
print('y_from: max(0,', box.bottom - self.h)
print('y_to: min', im_h - self.h, box.top)
print(y_from, y_to)
sys.stdout.flush()
crop = [x, y, x + self.w, y + self.h]
image = image.crop(crop)
new_boxes = []
new_labels = []
for box, label in zip(boxes, labels):
iom = RandomCrop._iom(crop,
[box.left, box.top, box.right, box.bottom])
# print(iom)
if iom < self.min_visibility:
continue
new_boxes.append(
BoundingBox(max(0, box.left - x), max(0, box.top - y),
min(self.w - 1, box.right - x),
min(self.h - 1, box.bottom - y), self.w, self.h,
label))
new_labels.append(label)
if len(new_boxes) == 0:
print(new_boxes)
box = orig
print('bbox:', box.left, box.top, box.right, box.bottom)
print('im_h: {}, im_w: {}'.format(im_h, im_w))
print('x:')
print('x_from; max(0,', box.right - self.w)
print('x_to: min', im_w - self.w, box.left)
print(x_from, x_to)
print('y:')
print('y_from: max(0,', box.bottom - self.h)
print('y_to: min', im_h - self.h, box.top)
print(y_from, y_to)
sys.stdout.flush()
return {'image': image, 'boxes': new_boxes, 'labels': new_labels}