def augmentation(self,
idx=None,
do_aug=True,
relative=True,
resize=True,
for_one_image=None,
show_img=False):
"""Create augmentation images from kitti.
:param idx: the index of kitti images ,
index is in the shape of [1, 3, 555, 1033...]
:param do_aug: if do_aug is False , then do nothing about images,and labels,
but corp to fixed size.
:param relative: the relative location of bounding boxes
:return: return a np array of images, and return a list about
labs are in the shape of [[class, left, top, button, right],..]
"""
if resize:
resize = random.choice(
self.cfg.TRAIN.MULTI_SIZE_RATIO) * self.cfg.TRAIN.IMG_SIZE
images, labels = self._augmenting(idx,
do_aug,
relative,
resize=resize,
image_for_aug=for_one_image)
if labels is not None: # if labels is not none ,then check the label in labels,whether the label is none.
for i, label in enumerate(labels):
time = 0
total_time = 50
while not label: # check every label, whether there is a label is empty.
print('no label at NO.%s, repeating %d/%d' %
(idx[i], time, total_time))
time += 1
re_aug_idx = [idx[i]] if time < total_time else [
random.choice(idx)
]
image, label = self._augmenting(
re_aug_idx,
do_aug,
relative,
resize=resize,
image_for_aug=for_one_image)
label = label[0]
if label:
images[i] = image[0]
labels[i] = label
if show_img:
_show_img(images, labels, show_img=True, cfg=self.cfg)
return images, labels
def show_tracker_OTB(img_path, gt_path):
img_list = glob.glob(img_path + '*.jpg')
lines = open(gt_path).readlines()
for img, line in zip(img_list, lines):
img = cv2.imread(img)
box = line.strip().split(',')
box[0] = int(box[0])
box[1] = int(box[1])
box[2] = box[0] + int(box[2])
box[3] = box[1] + int(box[3])
boxes = ['OTB_GT']
boxes.append(box)
_show_img(img, [boxes], show_time=200)
def test_backbone(self, test_picture_path):
"""Test."""
# prepare paramertas
img_raw = cv2.imread(test_picture_path)
if img_raw is None:
print('ERROR：no such a image')
if self.cfg.TEST.DO_AUG:
img_aug, _ = self.dataaug.augmentation(for_one_image=img_raw)
img_aug = img_aug[0]
elif self.cfg.TEST.RESIZE:
img_aug = cv2.resize(img_raw, (int(
self.cfg.TRAIN.IMG_SIZE[1]), int(self.cfg.TRAIN.IMG_SIZE[0])))
else:
img_aug = img_raw
img_in = torch.from_numpy(img_aug).unsqueeze(0).type(
torch.FloatTensor).to(self.cfg.TRAIN.DEVICE)
img_raw = torch.from_numpy(img_raw).unsqueeze(0).type(
torch.FloatTensor)
img_in = img_in.permute([
0,
3,
1,
2,
])
img_in = img_in / 127.5 - 1.
predict = self.Model.forward(input_x=img_in, is_training=False)
labels_pre = self.parsepredict._parse_predict(predict)
return _show_img(img_raw,
labels_pre,
img_in=img_in[0],
pic_path=test_picture_path,
cfg=self.cfg)
def test_backbone(self, test_picture_path):
"""Test."""
# prepare paramertas
if self.args.checkpoint:
model_path = self.args.checkpoint
else:
model_path = self.cfg.PATH.TEST_WEIGHT_PATH
self.Model.load_state_dict(torch.load(model_path))
self.Model = self.Model.cuda()
self.Model.eval()
img_raw = cv2.imread(test_picture_path)
if img_raw is None:
print('ERROR：no such a image')
if self.cfg.TEST.DO_AUG:
img_aug, _ = self.dataaug.augmentation(for_one_image=img_raw,
do_aug=self.cfg.TEST.DO_AUG,
resize=self.cfg.TEST.RESIZE)
img_aug = img_aug[0]
elif self.cfg.TEST.RESIZE:
img_aug = cv2.resize(img_raw, (int(
self.cfg.TRAIN.IMG_SIZE[1]), int(self.cfg.TRAIN.IMG_SIZE[0])))
else:
img_aug = img_raw
img_in = torch.from_numpy(img_aug).unsqueeze(0).type(
torch.FloatTensor).cuda()
img_raw = torch.from_numpy(img_raw).unsqueeze(0).type(
torch.FloatTensor)
predict = self.Model.forward(img_in)
labels_pre = self.parsepredict._parse_predict(predict)
return _show_img(img_raw,
labels_pre,
img_in=img_in[0],
pic_path=test_picture_path,
cfg=self.cfg)
def test_set(self):
set_path = '../tmp/checkpoint/apollo_lg_1_test_set'
test_set = torch.load(set_path)
do_aug = False
resize = True
for img_idx in test_set:
if os.path.isfile(self.cfg.IMGPATH + '%06d.png' % img_idx):
file_name = '%06d.png' % img_idx
else:
file_name = '%06d.jpg' % img_idx
imagepath = self.cfg.IMGPATH + file_name
# label_path = cfg.LABPATH + '%06d.txt' % img_idx
image1 = self.test_backbone(imagepath, do_aug, resize)
# imgs, labels = self.augmentation([img_idx], do_aug=False, resize=False)
image2 = _show_img([img_idx], do_aug, resize)
image_cat = np.vstack((image1, image2))
cv2.imshow('img', image_cat)
cv2.waitKey()
def _show_imgaug(self,
idx,
do_aug=True,
resize=True,
relative=False,
show_img=True):
"""
Show the images with data augmentation.
:param idx: idx of images from kitti.
:return: show the images.
"""
images, labels = self.augmentation(idx,
do_aug=do_aug,
relative=relative,
resize=resize)
images = torch.Tensor(images)
LOG.debug(labels)
return _show_img(images, labels, show_img=show_img, cfg=self.cfg)
def get_data_by_idx(self, idx_store, index_from, index_to, is_training):
'''
:param idx_store:
:param index_from:
:param index_to:
:return: imags: torch.Float32, relative labels:[[cls, x1, y1, x2, y2],[...],...]
'''
data = (None, None)
if self.one_test:
if self.one_name:
idx = self.one_name
else:
idx = idx_store[1:2]
else:
idx = idx_store[index_from:index_to]
if not idx:
print('error, no IDX in loader_img.py')
exit()
imgs, labels = self._read_datas(idx)
do_aug = self.cfg.TRAIN.DO_AUG if is_training else self.cfg.TEST.DO_AUG
if do_aug:
imgs, labels = self.dataaug.augmentation((imgs, labels))
if self.cfg.TRAIN.RESIZE:
size = random.choice(
self.cfg.TRAIN.MULTI_SIZE_RATIO) * self.cfg.TRAIN.IMG_SIZE
resized_imgs = []
resized_labels = []
for i, img_i in enumerate(imgs):
img_i_size = img_i.shape
resized_imgs.append(cv2.resize(img_i, (size[1], size[0])))
labs_i = labels[i]
if self.cfg.TRAIN.RELATIVE_LABELS:
label_i = [[
lab[0], lab[1] / img_i_size[1], lab[2] / img_i_size[0],
lab[3] / img_i_size[1], lab[4] / img_i_size[0]
] for lab in labs_i]
else:
label_i = [[
lab[0], lab[1] / img_i_size[1] * size[1],
lab[2] / img_i_size[0] * size[0],
lab[3] / img_i_size[1] * size[1],
lab[4] / img_i_size[0] * size[0]
] for lab in labs_i]
resized_labels.append(label_i)
imgs = resized_imgs
labels = resized_labels
if self.cfg.TRAIN.SHOW_INPUT:
_show_img(imgs, labels, show_img=True, cfg=self.cfg)
imgs = torch.Tensor(np.array(imgs))
imgs = imgs.permute([
0,
3,
1,
2,
])
imgs = imgs / 127.5 - 1.
imgs = imgs.to(self.cfg.TRAIN.DEVICE)
data = (imgs, labels) #
return data