def pull_item(self, index):
while True:
img_id = self.ids[index]
img_path = self._imgpath % img_id
target = ET.parse(self._annopath % img_id).getroot()
img = Image.open(img_path)
if img.mode == 'L':
img = img.convert('RGB')
width, height = img.size
if self.target_transform is not None:
target = self.target_transform(target, width, height)
bbox_labels = target
target = np.array(target)
if target.ndim != 2:
index = random.randrange(0, len(self.ids))
continue
img, sample_labels = preprocess(img, bbox_labels, self.mode,
img_path)
sample_labels = np.array(sample_labels)
if len(sample_labels) > 0:
target = np.hstack(
(sample_labels[:, 1:], sample_labels[:, 0][:, np.newaxis]))
assert (target[:, 2] > target[:, 0]).any()
assert (target[:, 3] > target[:, 1]).any()
break
else:
index = random.randrange(0, len(self.ids))
return torch.from_numpy(img), target
def pull_item(self, index):
while True:
image_path = self.fnames[index]
img = Image.open(image_path)
if img.mode == 'L':
img = img.convert('RGB')
im_width, im_height = img.size
boxes = self.annotransform(np.array(self.boxes[index]), im_width,
im_height)
label = np.array(self.labels[index])
bbox_labels = np.hstack((label[:, np.newaxis], boxes)).tolist()
img, sample_labels = preprocess(img, bbox_labels, self.mode,
image_path)
sample_labels = np.array(sample_labels)
if len(sample_labels) > 0:
target = np.hstack(
(sample_labels[:, 1:], sample_labels[:, 0][:, np.newaxis]))
assert (target[:, 2] > target[:, 0]).any()
assert (target[:, 3] > target[:, 1]).any()
break
else:
index = random.randrange(0, self.num_samples)
if self.mono_mode == True:
im = 0.299 * img[0] + 0.587 * img[1] + 0.114 * img[2]
return torch.from_numpy(np.expand_dims(
im, axis=0)), target, im_height, im_width
return torch.from_numpy(img), target, im_height, im_width
def pull_item(self, index):
while True:
# image_path = self.fnames[index]
image_path = os.path.join(self.root, self.fnames[index])
img = Image.open(image_path)
if img.mode == 'L':
img = img.convert('RGB')
im_width, im_height = img.size
boxes = self.annotransform(np.array(self.boxes[index]), im_width,
im_height)
label = np.array(self.labels[index])
bbox_labels = np.hstack((label[:, np.newaxis], boxes)).tolist()
img, sample_labels = preprocess(img, bbox_labels, self.mode,
image_path)
sample_labels = np.array(sample_labels)
if len(sample_labels) > 0:
face_target = np.hstack(
(sample_labels[:, 1:], sample_labels[:, 0][:, np.newaxis]))
assert (face_target[:, 2] > face_target[:, 0]).any()
assert (face_target[:, 3] > face_target[:, 1]).any()
# img = img.astype(np.float32)
face_box = face_target[:, :-1]
head_box = self.expand_bboxes(face_box)
head_target = np.hstack(
(head_box, face_target[:, -1][:, np.newaxis]))
break
else:
index = random.randrange(0, self.num_samples)
return torch.from_numpy(img), face_target, head_target
def pull_item(self, index):
while True:
image_path = self.fnames[index]
img = np.asarray(cv2.imread(image_path, cv2.IMREAD_COLOR), dtype=np.uint8)[...,::-1] # BGR -> RGB
boxes = np.array(self.boxes[index], dtype=np.int32)
if self.mode=='train' and cfg.apply_distort:
auged_d = spaaug(image=img, bboxes=boxes, category_id=[0]*len(boxes))
img, boxes = auged_d['image'], np.asarray(auged_d['bboxes'], dtype=np.float32)
im_height, im_width = img.shape[:2]
if len(boxes)>0: # if has bounding boxes
boxes = np.round(np.asarray(boxes, dtype=np.float32)).astype(np.int32)
boxes[:,[0,2]] = boxes[:,[0,2]].clip(0, im_width-1)
boxes[:,[1,3]] = boxes[:,[1,3]].clip(0, im_height-1)
boxes = boxes[(boxes[:,0]<boxes[:,2]) & (boxes[:,1]<boxes[:,3])] # delete invalid bbox
boxes = boxes.astype(np.float32)
if len(boxes)>0:
boxes = self.annotransform(boxes, im_width, im_height)
bbox_labels = np.pad(boxes, ((0,0),(1,0)), 'constant', constant_values=1.0).tolist()
img, sample_labels = preprocess(
img, bbox_labels, self.mode, image_path)
else:
sample_labels = []
sample_labels = np.array(sample_labels)
if len(sample_labels) > 0:
face_target = np.hstack(
(sample_labels[:, 1:], sample_labels[:, 0][:, np.newaxis]))
assert (face_target[:, 2] > face_target[:, 0]).all()
assert (face_target[:, 3] > face_target[:, 1]).all()
#img = img.astype(np.float32)
face_box = face_target[:, :-1]
head_box = self.expand_bboxes(face_box)
head_target = np.hstack((head_box, face_target[
:, -1][:, np.newaxis]))
break
else:
index = random.randrange(0, self.num_samples)
img_torch = torch.from_numpy(img)
'''
img = img[::-1]
img += cfg.img_mean
img = img[::-1]
img = np.transpose(img, (1,2,0)).astype(np.uint8)
img = Image.fromarray(img)
draw = ImageDraw.Draw(img)
w,h = img.size
for bbox in sample_labels:
bbox = (bbox[1:] * np.array([w, h, w, h])).tolist()
draw.rectangle(bbox,outline='red')
img.save('{:d}.jpg'.format(index))
'''
return img_torch, face_target, head_target
def pull_item(self, index):
while True:
image_path = self.fnames[index]
img = Image.open(image_path)
if img.mode == 'L':
img = img.convert('RGB')
boxes = self.boxes[index]
# Apply aug
img = np.array(img) # (h, w, c)
if self.mode == 'train' and cfg.apply_distort:
img, boxes = self.augmentation.ultimate_aug(img, boxes)
img = Image.fromarray(img)
im_width, im_height = img.size
boxes = self.annotransform(np.array(boxes), im_width, im_height)
label = np.array(self.labels[index][:boxes.shape[0]])
bbox_labels = np.hstack((label[:, np.newaxis], boxes)).tolist()
img, sample_labels = preprocess(img, bbox_labels, self.mode,
image_path)
sample_labels = np.array(sample_labels)
if len(sample_labels) > 0:
target = np.hstack(
(sample_labels[:, 1:], sample_labels[:, 0][:, np.newaxis]))
assert (target[:, 2] > target[:, 0]).any()
assert (target[:, 3] > target[:, 1]).any()
break
else:
index = random.randrange(0, self.num_samples)
#img = Image.fromarray(img)
'''
draw = ImageDraw.Draw(img)
w,h = img.size
for bbox in sample_labels:
bbox = (bbox[1:] * np.array([w, h, w, h])).tolist()
draw.rectangle(bbox,outline='red')
img.save('image.jpg')
'''
return torch.from_numpy(img), target, im_height, im_width
def pull_item(self, index):
while True:
image_path = self.fnames[index]
img = Image.open(image_path)
if img.mode == 'L':
img = img.convert('RGB')
im_width, im_height = img.size
boxes = self.annotransform(
np.array(self.boxes[index]), im_width, im_height)
label = np.array(self.labels[index])
bbox_labels = np.hstack((label[:, np.newaxis], boxes)).tolist()
img, sample_labels = preprocess(
img, bbox_labels, self.mode, image_path)
sample_labels = np.array(sample_labels)
if len(sample_labels) > 0:
face_target = np.hstack(
(sample_labels[:, 1:], sample_labels[:, 0][:, np.newaxis]))
assert (face_target[:, 2] > face_target[:, 0]).any()
assert (face_target[:, 3] > face_target[:, 1]).any()
#img = img.astype(np.float32)
face_box = face_target[:, :-1]
head_box = self.expand_bboxes(face_box)
head_target = np.hstack((head_box, face_target[
:, -1][:, np.newaxis]))
break
else:
index = random.randrange(0, self.num_samples)
#img = Image.fromarray(img)
'''
draw = ImageDraw.Draw(img)
w,h = img.size
for bbox in sample_labels:
bbox = (bbox[1:] * np.array([w, h, w, h])).tolist()
draw.rectangle(bbox,outline='red')
img.save('image.jpg')
'''
return torch.from_numpy(img), face_target, head_target
def pull_item(self, index):
while True:
try:
image_path = self.fnames[index]
img = Image.open(image_path)
if img.mode == 'L':
img = img.convert('RGB')
im_width, im_height = img.size
boxes = self.annotransform(
np.array(self.boxes[index]), im_width, im_height)
label = np.array(self.labels[index])
bbox_labels = np.hstack((label[:, np.newaxis], boxes)).tolist()
img, sample_labels = preprocess(
img, bbox_labels, self.mode, image_path)
sample_labels = np.array(sample_labels)
if len(sample_labels) > 0:
target = np.hstack(
(sample_labels[:, 1:], sample_labels[:, 0][:, np.newaxis]))
assert (target[:, 2] > target[:, 0]).any()
assert (target[:, 3] > target[:, 1]).any()
break
else:
index = random.randrange(0, self.num_samples)
except Exception as e:
print('Error in WIDERDetection:', image_path, e)
index = random.randrange(0, self.num_samples)
#img = Image.fromarray(img)
'''
draw = ImageDraw.Draw(img)
w,h = img.size
for bbox in sample_labels:
bbox = (bbox[1:] * np.array([w, h, w, h])).tolist()
draw.rectangle(bbox,outline='red')
img.save('image.jpg')
'''
return torch.from_numpy(img), target, im_height, im_width