def imshow_segms(img_or_path,
segms=None,
labels=None,
scores=None,
score_threshold=0.0,
colors='red',
thickness=3,
show=True,
win_name='',
wait_time=0,
out_file=None,
return_img=False,
draw_contours=False):
grayscale_image = img_or_path.astype(np.float64)
max_value = np.max(grayscale_image)
min_value = np.min(grayscale_image)
grayscale_image = 255 * (grayscale_image - min_value) / (max_value - min_value)
grayscale_image = grayscale_image.astype(np.uint8)
grayscale_image = np.pad(grayscale_image, ((25, 25), (25, 25)), 'constant', constant_values = (0, 0))
# grayscale_image[grayscale_image < 128] = 0
ret, grayscale_image = cv2.threshold(grayscale_image, 128, 255, 0)
nLabels, labels, stats, centroids = cv2.connectedComponentsWithStats(grayscale_image.astype(np.uint8), connectivity=4)
det = []
mapper = []
for k in range(1, nLabels):
# size filtering
size = stats[k, cv2.CC_STAT_AREA]
if size < 10: continue
x, y = stats[k, cv2.CC_STAT_LEFT], stats[k, cv2.CC_STAT_TOP]
w, h = stats[k, cv2.CC_STAT_WIDTH], stats[k, cv2.CC_STAT_HEIGHT]
niter = int(np.sqrt(size * min(w, h) / (w * h)) * 2)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT,(1 + niter, 1 + niter))
grayscale_image = cv2.dilate(grayscale_image, kernel)
# make box
np_contours = np.roll(np.array(np.where(grayscale_image!=0)),1,axis=0).transpose().reshape(-1,2)
rect = cv2.minAreaRect(np_contours)
box = cv2.boxPoints(rect)
box = np.int0(box)
cv2.drawContours(grayscale_image, [box], 0, (0, 0, 255), 3)
x, y, w, h, theta = rect[0][0], rect[0][1], rect[1][0], rect[1][1], rect[2]
theta = theta * np.pi / 180.0
thetaobb = [x, y, w, h, theta]
print(thetaobb)
pointobb = thetaobb2pointobb(thetaobb)
print(pointobb)
imshow_rbboxes(grayscale_image, pointobb, win_name='demo')
def __hrsc_parse__(self, hrsc_label_file, hrsc_image_file):
objects = []
if self.groundtruth:
tree = ET.parse(hrsc_label_file)
root = tree.getroot()
objects = []
hrsc_object = root.find('HRSC_Objects')
for hrsc_sub_object in hrsc_object.findall('HRSC_Object'):
obj_struct = {}
xmin = float(hrsc_sub_object.find('box_xmin').text)
ymin = float(hrsc_sub_object.find('box_ymin').text)
xmax = float(hrsc_sub_object.find('box_xmax').text)
ymax = float(hrsc_sub_object.find('box_ymax').text)
bbox_w = xmax - xmin
bbox_h = ymax - ymin
cx = float(hrsc_sub_object.find('mbox_cx').text)
cy = float(hrsc_sub_object.find('mbox_cy').text)
rbbox_w = float(hrsc_sub_object.find('mbox_w').text)
rbbox_h = float(hrsc_sub_object.find('mbox_h').text)
angle = float(hrsc_sub_object.find('mbox_ang').text)
# angle = angle * 180.0 / np.pi
obj_struct['bbox'] = [xmin, ymin, bbox_w, bbox_h]
obj_struct['thetaobb'] = [cx, cy, rbbox_w, rbbox_h, angle]
obj_struct['segmentation'] = thetaobb2pointobb(
obj_struct['thetaobb'])
obj_struct['pointobb'] = pointobb_sort_function[
pointobb_sort_method](obj_struct['segmentation'])
obj_struct['keypoints'] = obj_struct['pointobb'][:]
for idx in [2, 5, 8, 11]:
obj_struct['keypoints'].insert(idx, 2)
obj_struct['hobb'] = thetaobb2hobb(
obj_struct['thetaobb'],
pointobb_sort_function[pointobb_sort_method])
obj_struct['label'] = 1
objects.append(obj_struct)
else:
obj_struct = {}
obj_struct['segmentation'] = [0, 0, 0, 0, 0, 0, 0, 0]
obj_struct['keypoint'] = [0, 0, 0, 0, 0, 0, 0, 0]
obj_struct['pointobb'] = [0, 0, 0, 0, 0, 0, 0, 0]
obj_struct['thetaobb'] = [0, 0, 0, 0, 0]
obj_struct['hobb'] = [0, 0, 0, 0, 0]
obj_struct['bbox'] = [0, 0, 0, 0]
obj_struct['label'] = 0
objects.append(obj_struct)
return objects
def imshow_rbboxes(img_or_path,
rbboxes,
labels=None,
scores=None,
score_threshold=0.0,
colors='red',
show_label=False,
show_score=False,
thickness=3,
show=True,
win_name='',
wait_time=0,
out_file=None,
return_img=False):
""" Draw oriented bounding boxes on image
Args:
img (str or ndarray): The image to be displayed.
rbboxes (list or ndarray): A ndarray of shape (N, 8)
labels (list or ndarray): A ndarray of shape (N, 1)
scores (list or ndarray): A ndarray of shape (N, 1)
"""
if is_str(img_or_path):
img = cv2.imread(img_or_path)
else:
img = img_or_path
if rbboxes == []:
return
if isinstance(rbboxes, list):
rbboxes = np.array(rbboxes)
if rbboxes.shape[1] == 5:
rbboxes_ = []
for rbbox in rbboxes:
rbboxes_.append(thetaobb2pointobb(rbbox))
rbboxes = np.array(rbboxes_)
if rbboxes.ndim == 1:
rbboxes = np.array([rbboxes])
if labels is None:
labels_vis = np.array(['ins'] * rbboxes.shape[0])
else:
labels_vis = np.array(labels)
if labels_vis.ndim == 0:
labels_vis = np.array([labels_vis])
if scores is None:
scores_vis = np.array([1.0] * rbboxes.shape[0])
else:
scores_vis = np.array(scores)
if scores_vis.ndim == 0:
scores_vis = np.array([scores_vis])
if labels is None:
colors = dict()
colors[colors] = color_val(colors)
else:
max_label = max(labels)
colors = [color_val(_) for _ in range(max_label + 1)]
for rbbox, label, score in zip(rbboxes, labels_vis, scores_vis):
if score < score_threshold:
continue
if len(rbbox) == 5:
rbbox = np.array(thetaobb2pointobb(rbbox))
rbbox = rbbox.astype(np.int32)
cx = np.mean(rbbox[::2])
cy = np.mean(rbbox[1::2])
current_color = colors[label]
for idx in range(-1, 3, 1):
cv2.line(
img, (int(rbbox[idx * 2]), int(rbbox[idx * 2 + 1])),
(int(rbbox[(idx + 1) * 2]), int(rbbox[(idx + 1) * 2 + 1])),
current_color,
thickness=thickness)
if show_label:
cv2.putText(img,
label, (cx, cy),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
fontScale=1.0,
color=current_color,
thickness=2,
lineType=8)
if show_score:
cv2.putText(img,
"{:.2f}".format(score), (cx, cy),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
fontScale=1.0,
color=current_color,
thickness=2,
lineType=8)
if show:
cv2.namedWindow(win_name, cv2.WINDOW_NORMAL)
cv2.imshow(win_name, img)
cv2.waitKey(wait_time)
if out_file is not None:
dir_name = osp.abspath(osp.dirname(out_file))
mkdir_or_exist(dir_name)
cv2.imwrite(out_file, img)
if return_img:
return img
],
[
57.678192138671875, 259.0009765625, 83.1702880859375,
179.4698944091797, -0.4243928624890527
],
[
264.6916809082031, 741.2076416015625, 82.70415496826172,
145.08529663085938, -0.42366891543622837
],
[
373.6033630371094, 713.7134399414062, 81.72265625,
179.64755249023438, -0.4245613407175139
]]
for thetaobb in thetaobbs:
pointobb = thetaobb2pointobb(thetaobb)
pointobbs.append(pointobb)
pseudomasks = np.zeros((height, width), dtype=np.int32)
for pointobb in pointobbs:
transformed, mask_location = pointobb2pseudomask(pointobb,
anchor_image,
host_height=height,
host_width=width)
# show_grayscale_as_heatmap(transformed)
transformed = transformed.astype(np.int32)
pseudomasks[mask_location[1]:mask_location[3],
mask_location[0]:mask_location[2]] = np.where(
transformed >
pseudomasks[mask_location[1]:mask_location[3],
mask_location[0]:mask_location[2]],