def get_report_df(filesdir, lengthValue):
filepaths = NumDensityReporter.scanAllFiles(filesdir)
report_df = pd.DataFrame(
columns=['dir', 'xmlfile', 'index', 'diameter'])
for i_xml, xmlPath in enumerate(filepaths):
*dir_, _ = xmlPath.split('_')
dir_ = os.path.basename("_".join(dir_)) if isinstance(
dir_, (list, tuple)) else os.path.basename(dir_)
xmlfilename = os.path.basename(xmlPath)
if os.path.isfile(xmlPath) is False:
continue
tVocParseReader = PascalVocReader(xmlPath)
shapes = tVocParseReader.getShapes()
i_label = 0
i_track = 0
i_easytrack = 0
for shape in shapes:
shapeType, label, points, _, _, difficult = shape
if shapeType == shapeTypes.ellipse:
diameter = getDiameter(points) * lengthValue
report_df = report_df.append([{
'dir': dir_,
'xmlfile': xmlfilename,
'index': i_label,
'diameter': diameter
}])
i_label += 1
return report_df
def get_meta(image_path=None, xml_dir=None):
basename = os.path.basename(os.path.splitext(image_path)[0])
meta_path = os.path.join(xml_dir, basename + XML_EXT)
if os.path.isfile(meta_path):
reader = PascalVocReader(meta_path)
reader.parseXML()
return reader
else:
return None
def __init__(self, filename=None):
if filename is not None:
reader = PascalVocReader(filename)
self.shapes = reader.getShapes()
self.imagePath = reader.imagePath
self.imageData = read(self.imagePath, None)
self.verified = reader.verified
else:
self.shapes = ()
self.imagePath = None
self.imageData = None
self.verified = False
def loadPascalXMLByFilename(self, xmlPath):
if self.filePath is None:
return
if os.path.isfile(xmlPath) is False:
return
tVocParseReader = PascalVocReader(xmlPath)
shapes = tVocParseReader.getShapes()
self.loadLabels(shapes)
self.canvas.verified = tVocParseReader.verified
def get_image(filename,label_num_dic = None):
tVocParseReader = PascalVocReader(filename)
raw_shapes = tVocParseReader.getShapes()
print raw_shapes
def format_shape(s):
label,points,aa,bb,c = s
return dict(label=unicode(label),
points=[(int(p[0]), int(p[1])) for p in points])
shapes = [format_shape(shape) for shape in raw_shapes]
image_size = tVocParseReader.get_img_size()
result_path = '/mask'+filename.split('/')[1].split('.')[0]+'.png'
mask_writer = save_mask_image.label_mask_writer(label_num_dic, result_path, image_size[0],
image_size[1])
mask_writer.save_mask_image(shapes)
def generate_mask_png(label_num_dic, xml_file_name, output_path):
if os.path.exists(output_path) is False:
os.makedirs(output_path)
tVocParseReader = PascalVocReader(xml_file_name)
shapes = tVocParseReader.getShapes()
image_size = tVocParseReader.get_img_size()
num = 1
no_ext_name = os.path.splitext(xml_file_name)
for shape in shapes:
final_path = no_ext_name + u'_' + unicode(num) + '.png'
mask_writer = label_mask_writer(label_num_dic, final_path,
image_size[0], image_size[1])
mask_writer.save_mask_image(shapes)
num = num + 1
def get_label_info(self, dir_path):
labels_path = os.path.join(dir_path, 'labels')
labels_exist_path = labels_path if os.path.exists(
labels_path) else dir_path
self.label_to_num = {}
for file_name in tqdm(os.listdir(labels_exist_path), "check label"):
if os.path.splitext(file_name)[1] != XML_EXT:
continue
xml_path = os.path.join(labels_exist_path, file_name)
reader = PascalVocReader(xml_path)
for shape in reader.shapes:
if shape[0] not in self.label_to_num:
self.label_to_num.setdefault(shape[0], 1)
else:
self.label_to_num[shape[0]] += 1
list_of_dataset = os.listdir(dir_path)
self.label_to_num['origin_image_num'] = 0
self.label_to_num['labeled_image_num'] = 0
for each_item in list_of_dataset:
if os.path.splitext(each_item)[1] == '.jpg':
self.label_to_num['origin_image_num'] += 1
else:
self.check_label_exist(os.path.join(dir_path, each_item))
QMessageBox.information(None, 'info', str(self.label_to_num))
def get_image(filename, label_num_dic=None):
tVocParseReader = PascalVocReader(filename)
raw_shapes = tVocParseReader.getShapes()
print raw_shapes
def format_shape(s):
label, points, aa, bb, c = s
return dict(label=unicode(label),
points=[(int(p[0]), int(p[1])) for p in points])
shapes = [format_shape(shape) for shape in raw_shapes]
image_size = tVocParseReader.get_img_size()
result_path = '/mask' + filename.split('/')[1].split('.')[0] + '.png'
mask_writer = save_mask_image.label_mask_writer(label_num_dic, result_path,
image_size[0],
image_size[1])
mask_writer.save_mask_image(shapes)
def get_easy_track_report_df(dir, lengthValue, labelHist):
easy_track_df = pd.DataFrame(columns=[
'dir', 'xmlfilename', 'index', 'label', 'shapeType', 'points',
'difficult'
])
filepaths = NumDensityReporter.scanAllFiles(dir)
for i_xml, xmlPath in enumerate(filepaths):
*dir_, _ = xmlPath.split('_')
dir_ = os.path.basename("_".join(dir_)) if isinstance(
dir_, (list, tuple)) else os.path.basename(dir_)
xmlfilename = os.path.basename(xmlPath)
if os.path.isfile(xmlPath) is False:
continue
tVocParseReader = PascalVocReader(xmlPath)
shapes = tVocParseReader.getShapes()
i_easytrack = 0
for shape in shapes:
shapeType, label, points, _, _, difficult = shape
if label not in labelHist:
easy_track_df = easy_track_df.append([{
'dir':
dir_,
# 'subdir':subdir,
'xmlfilename':
xmlfilename,
'index':
i_easytrack,
'label':
label,
'shapeType':
shapeType,
'points':
points,
'difficult':
difficult
}])
i_easytrack += 1
return easy_track_df
def visualize(vis_params,
logger,
img_path,
csv_path,
class_dict,
init_frame_id=0,
n_frames=0,
request_roi=None,
generator_mode=0,
enable_masks=0,
label='',
only_boxes=0,
crop_size=()):
"""
:param vis_params:
:param logger:
:param img_path:
:param csv_path:
:param class_dict:
:param init_frame_id:
:param n_frames:
:param request_roi:
:param generator_mode:
:return:
"""
global _pause, _quit
save_fname_templ = os.path.splitext(os.path.basename(img_path))[0]
# csv_path = os.path.join(img_path, 'annotations.csv')
df = pd.read_csv(csv_path)
frames_reader = get_frames_reader(img_path, save_as_bin=False)
if request_roi is not None:
frames_reader.setROI(request_roi)
class_labels = dict((v, k) for k, v in class_dict.items())
if generator_mode:
vis_params.show = 0
vis_params.save = 0
visualizer = Visualizer(vis_params, logger, class_labels)
init_frame = frames_reader.get_frame(init_frame_id)
height, width, _ = init_frame.shape
frame_size = width, height
visualizer.initialize(save_fname_templ, frame_size, _pause)
if n_frames <= 0:
n_frames = frames_reader.num_frames
print('Reading {:d} images from {:s}...'.format(n_frames, img_path))
for frame_id in range(init_frame_id, n_frames):
try:
curr_frame = frames_reader.get_frame(frame_id)
except IOError as e:
print('{}'.format(e))
break
if only_boxes:
curr_frame = np.zeros_like(curr_frame)
file_path = frames_reader.get_file_path()
if file_path is None:
print('Visualization is only supported on image sequence data')
return
filename = os.path.basename(file_path)
multiple_instance = df.loc[df['filename'] == filename]
# Total # of object instances in a file
n_bboxes = len(multiple_instance.index)
# Remove from df (avoids duplication)
df = df.drop(multiple_instance.index[:n_bboxes])
frame_data = []
masks = []
generic_target_id = -1
if enable_masks:
filename = os.path.basename(file_path)
xml_path = os.path.join(img_path, 'annotations',
os.path.splitext(filename)[0] + '.xml')
if not os.path.isfile(xml_path):
print('{} :: annotations xml file not found: {}'.format(
filename, xml_path))
continue
xml_reader = PascalVocReader(xml_path)
shapes = xml_reader.getShapes()
n_shapes = len(shapes)
if n_shapes != n_bboxes:
raise IOError(
'Mismatch between n_bboxes in xml: {} and csv: {}'.format(
n_shapes, n_bboxes))
for box_id in range(n_bboxes):
bbox = multiple_instance.iloc[box_id]
try:
target_id = bbox['target_id']
except KeyError:
target_id = generic_target_id
generic_target_id -= 1
xmin = bbox.loc['xmin']
ymin = bbox.loc['ymin']
xmax = bbox.loc['xmax']
ymax = bbox.loc['ymax']
class_name = bbox.loc['class']
try:
class_id = class_dict[str(class_name)]
except KeyError:
print('Ignoring annotation with invalid class: {}'.format(
class_name))
continue
width = xmax - xmin
height = ymax - ymin
curr_frame_data = [
frame_id, target_id, xmin, ymin, width, height, class_id
]
if enable_masks:
mask = shapes[box_id][-2]
if mask is not None:
_contour_pts = Shape.getContourPts(mask)
masks.append(_contour_pts)
frame_data.append(curr_frame_data)
frame_data = np.asarray(frame_data)
res = visualizer.update(frame_id, curr_frame, frame_data, masks, label,
crop_size)
if generator_mode:
yield res
# elif not res:
# break
_quit = visualizer._quit
_pause = visualizer._pause
visualizer.close()
frames_reader.close()
printname=Path(name).stem
xmlfilepath=PurePath(inputpath, filexml)
jpgfilepath=PurePath(outputpath, filename)
pngfilepath=PurePath(outputpath, filepng)
print("XML file %s" % xmlfilepath)
reader = PascalVocReader(xmlfilepath)
shapes = reader.getShapes()
print(reader.verified)
print(shapes)
print(len(shapes))
#[('salmon', [(1, 262), (337, 262), (337, 588), (1, 588)], None, None, True)]
# Scale font to picture but anything smaller than 0.55 is not readable
font_scale=max(float(cols/1000),0.55)
#fontsize=float(cols/1000)
text_offset_x=int(cols/40)
text_offset_y=int(rows/40) + 10
#linesize=1
font = cv.FONT_HERSHEY_PLAIN
#double scale = 0.4;
filenames = sorted(glob.glob(IN_DIR + "/*.png"))
count = 0
for filename in filenames:
xmlName = filename[:-4] + ".xml"
basename = os.path.basename(filename)
dirname = os.path.basename(os.path.dirname(filename))
if (not os.path.exists(TRAIN_OUT_DIR + dirname)):
os.makedirs(TRAIN_OUT_DIR + dirname)
if (not os.path.exists(VALID_OUT_DIR + dirname)):
os.makedirs(VALID_OUT_DIR + dirname)
src = cv2.imread(filename)
xmlReader = PascalVocReader(xmlName)
shapes = xmlReader.getShapes()
xmin = src.shape[1]
ymin = src.shape[0]
xmax = 0
ymax = 0
for i in range(len(shapes)):
if (xmin > shapes[i][1][0][0]):
xmin = shapes[i][1][0][0]
if (ymin > shapes[i][1][0][1]):
ymin = shapes[i][1][0][1]
if (xmax < shapes[i][1][2][0]):
xmax = shapes[i][1][2][0]
if (ymax < shapes[i][1][2][1]):
def loadPascalXMLByFilename(xmlPath):
tVocParseReader = PascalVocReader(xmlPath)
shapes = tVocParseReader.getShapes()
return shapes
def run(self):
xmls = []
for index, (dirpath, dirnames,
filenames) in enumerate(os.walk(self.dir)):
xmls.extend(
list(
map(
lambda s: os.path.join(dirpath, s),
filter(lambda s: True
if s.endswith(".xml") else False, filenames))))
easy_track_df = pd.DataFrame()
for xml in xmls:
tVocParseReader = PascalVocReader(xml)
shapes = tVocParseReader.getShapes()
xmlfilename = os.path.basename(xml)
*dir_, _ = xml.split('_')
dir_ = os.path.basename("_".join(dir_)) if isinstance(
dir_, (list, tuple)) else os.path.basename(dir_)
i_easytrack = 0
for shape in shapes:
shapeType, label, points, _, _, difficult = shape
if label not in self.labelHist:
easy_track_df = easy_track_df.append(
[{
'subdir': dir_,
'xmlfilename': xmlfilename,
'index': i_easytrack,
'label': label,
'shapeType': shapeType,
'points': points,
'difficult': difficult
}],
ignore_index=True)
i_easytrack += 1
easy_track_df.sort_values(
by=["subdir", "xmlfilename", "label"]).reindex(
range(len(easy_track_df)))
ShapeObj = namedtuple('ShapeObj', ['subdir', 'label', 'shape'])
shape_obj_list = []
shape_obj_dict_list = []
subdirs = sorted(list(easy_track_df.subdir.value_counts().keys()))
for i_subdir, subdir in enumerate(subdirs):
tmp_sub_df = easy_track_df[easy_track_df["subdir"] ==
subdir].sort_values(by="label")
shape_obj_list_tmp = []
for i, row in tmp_sub_df.iterrows():
subdir = row.subdir
xmlfilename = row.xmlfilename
shapeType = row.shapeType
label = row.label.lower().strip()
points = row.points
difficult = row.difficult
shapeFac = shapeFactory()
shapeFac.setType(shapeType)
shape = shapeFac.getShape()
shape.points = [QPointF(*xy) for xy in points]
shape.label = label
shape.difficult = difficult
shape.close()
shape_obj = ShapeObj(subdir, label, shape)
shape_obj_list_tmp.append(shape_obj)
shape_obj_dict = OrderedDict()
for shape in shape_obj_list_tmp:
just_found_one = False
key = " "
if " " not in shape.label:
key = shape.label
shape_obj_dict[key] = []
just_found_one = True
if shape.label.split(" ", 1)[0] in shape_obj_dict.keys():
if not just_found_one:
shape_obj_dict[shape.label.split(" ",
1)[0]].append(shape)
# 去除字典中小于1个的情况
to_del = []
for key, value in shape_obj_dict.items():
if len(value) <= 1:
to_del.append(key)
for key in to_del:
shape_obj_dict.pop(key)
shape_obj_list.extend(shape_obj_list_tmp)
shape_obj_dict_list.append(shape_obj_dict)
broken_info_dict = defaultdict(list)
for shape_obj_dict_tmp in shape_obj_dict_list:
for _, shape_objs in shape_obj_dict_tmp.items():
n_break = len(shape_objs)
diameters = tuple(shape_obj.shape.getDiameter()
for shape_obj in shape_objs)
if n_break in broken_info_dict.keys():
broken_info_dict[n_break].append(diameters)
else:
broken_info_dict[n_break] = [diameters]
def genSheetname(s):
return "broken {}".format(s)
def genColname(i):
return "diameter {}".format(i + 1)
df_dict = dict()
for key, value in broken_info_dict.items():
columns = [genColname(i) for i in range(key)] # range(len(value))
df = pd.DataFrame(value, columns=columns)
df['diameter of mother drop'] = reduce(
lambda x, y: (x**3 + y**3)**(1 / 3),
[df[col] for col in columns])
columns_frac = [col + " frac" for col in columns]
for col in columns_frac:
df[col] = df[
col[:-len(" frac")]]**3 / df['diameter of mother drop']**3
df_dict[key] = df
summary_df_data = []
for i_df, df in df_dict.items():
columns = [genColname(i) for i in range(i_df)]
for i, row in df.iterrows():
drops = list(row[columns])
assert (len(drops) == i_df)
drops.sort()
mother_drops = [0 for _ in range(len(drops))]
mother_drops[0] = drops[0]
for i in range(len(drops) - 1):
drop_1 = mother_drops[i]
drop_2 = drops[i + 1]
mother_drops[i + 1] = (drop_1**3 + drop_2**3)**(1 / 3)
drop_1_frac = drop_1**3 / mother_drops[i + 1]**3
drop_2_frac = drop_2**3 / mother_drops[i + 1]**3
n_breakage = i_df
summary_df_data.append(
(drop_1, drop_2, mother_drops[i + 1], drop_1_frac,
drop_2_frac, n_breakage))
summary_df = pd.DataFrame(summary_df_data,
columns=[
'drop_1', 'drop_2', 'mothoer_drop',
'drop_1_frac', "drop_2_frac",
"n_breakage"
])
filename = os.path.join(
os.path.dirname(self.dir),
"summary-{}.xlsx".format(os.path.basename(self.dir)))
writer = pd.ExcelWriter(filename)
for i_df, df in df_dict.items():
df.to_excel(writer, genSheetname(i_df))
summary_df.to_excel(writer, "summary")
writer.save()
text = "\n".join([
"# {i}-elem break: {num}".format(
i=str(i).rjust(2), num=str(len(broken_info_dict[i])).rjust(3))
for i in range(2, 11)
])
text += "\n" + "# total: {total}".format(total=len(summary_df))
self.finished.emit(text)
def saveMasks(seq_path,
xml_path,
out_mask_size,
out_border,
fixed_ar,
save_raw_mask,
show_img,
out_root_path='',
save_test=1,
save_train=1,
frames_reader=None,
masks_per_seq=0,
enable_out_suffix=1,
train_fnames=None,
test_fnames=None,
map_to_bbox=0,
out_img_dir='',
enable_xml_annotations=0,
allow_skipping_images=0):
global _pause, _exit
if not xml_path or not os.path.isdir(xml_path):
raise IOError(
'Folder containing the loaded boxes does not exist: {}'.format(
xml_path))
files = glob.glob(os.path.join(xml_path, '*.xml'))
n_files = len(files)
if n_files == 0:
raise IOError('No loaded boxes found')
if frames_reader is None:
frames_reader = get_frames_reader(seq_path, save_as_bin=False)
min_dim = max_dim = 0
out_w, out_h = out_mask_size
print('out_mask_size: {}'.format(out_mask_size))
if out_w == -1 and out_h == -1:
out_w = out_h = 0
if out_w == -1:
max_dim = out_h
elif out_h == -1:
min_dim = out_w
if fixed_ar:
print('Using fixed aspect ratio: {}'.format(fixed_ar))
print('out_border: {}'.format(out_border))
def getint(fn):
basename = os.path.basename(fn)
num = re.sub("\D", "", basename)
try:
return int(num)
except:
return 0
if len(files) > 0:
files = sorted(files, key=getint)
print('Loading annotations from {:d} files'.format(n_files))
file_id = 0
n_boxes = 0
seq_root_dir = os.path.dirname(seq_path)
seq_name = os.path.basename(seq_path)
if not out_root_path:
out_root_path = os.path.join(seq_root_dir, 'masks')
if not enable_out_suffix:
out_seq_name = seq_name
else:
if map_to_bbox:
out_seq_name = '{}_mapped'.format(seq_name)
else:
out_seq_name = '{}_{}x{}'.format(seq_name, out_w, out_h)
if fixed_ar:
out_seq_name = '{}_ar_{}'.format(out_seq_name, fixed_ar)
else:
out_seq_name = '{}_{}'.format(out_seq_name, out_border)
out_seq_name = out_seq_name.replace('.', 'p')
train_root_path = os.path.join(out_root_path, out_seq_name)
if not save_test and not save_train:
raise AssertionError('Either save_test or save_train must be on')
# print('Saving output sequences to {}'.format(out_root_path))
if save_train:
out_img_root_path = train_root_path
if out_img_dir:
out_img_root_path = os.path.join(out_img_root_path, out_img_dir)
out_mask_root_path = os.path.join(train_root_path, 'labels')
print('Saving training mask sequence to {}'.format(train_root_path))
if not os.path.isdir(out_img_root_path):
os.makedirs(out_img_root_path)
if not os.path.isdir(out_mask_root_path):
os.makedirs(out_mask_root_path)
if enable_xml_annotations:
out_xml_path = os.path.join(out_img_root_path, 'annotations')
print('Saving xml_annotations to {}'.format(out_xml_path))
if not os.path.isdir(out_xml_path):
os.makedirs(out_xml_path)
if save_test:
out_test_seq_name = out_seq_name + '_test'
test_img_root_path = os.path.join(out_root_path, out_test_seq_name)
print('Saving unlabeled testing mask sequence to {}'.format(
test_img_root_path))
if not os.path.isdir(test_img_root_path):
os.makedirs(test_img_root_path)
win_name = 'patch and mask'
disable_resizing = 0
scale_x = scale_y = 1.0
if out_w == 0 and out_h == 0:
print('Resizing disabled')
disable_resizing = 1
csv_raw = []
test_csv_raw = []
n_files = len(files)
if save_raw_mask:
print('Saving raw labels')
mask_pix_val = (1, 1, 1)
else:
mask_pix_val = (255, 255, 255)
n_masks = 0
_train_fnames = []
_test_fnames = []
_exit_seq = 0
disp_img = None
for file_id, file in enumerate(files):
xml_reader = PascalVocReader(file)
filename = os.path.basename(xml_reader.filename)
filename_no_ext = os.path.splitext(filename)[0]
# file_id = int(re.sub("\D", "", filename))
# print('filename: {}'.format(filename))
# print('file_id: {}'.format(file_id))
img = frames_reader.get_frame_by_name(filename, convert_to_rgb=0)
if img is None:
print('image {} could not be read'.format(filename))
continue
img_h, img_w = img.shape[:2]
mask_img = None
shapes = xml_reader.getShapes()
n_shapes = len(shapes)
if n_shapes > 1:
print('{} boxes found for {} in {}'.format(n_shapes, filename,
file))
obj_id = 0
img_written = 0
for shape in shapes:
label, points, _, _, difficult, bbox_source, id_number, score, mask, mask_img = shape
if not mask:
if not save_test:
continue
xmin, ymin = points[0]
xmax, ymax = points[2]
img_root_path = test_img_root_path
else:
if not save_train:
continue
mask_pts_list = Shape.getContourPts(mask, verbose=0)
mask_pts = np.asarray(mask_pts_list)
xmin, ymin = np.min(mask_pts, axis=0).astype(np.int32)
xmax, ymax = np.max(mask_pts, axis=0).astype(np.int32)
img_root_path = out_img_root_path
if fixed_ar:
w, h = xmax - xmin, ymax - ymin
src_ar = float(w) / float(h)
if fixed_ar > src_ar:
border_x = int((h * fixed_ar - w) / 2.0)
border_y = 0
else:
border_y = int((w / fixed_ar - h) / 2.0)
border_x = 0
else:
border_x = border_y = out_border
# start_row, start_col = max(0, ymin - border_y), max(0, xmin - border_x)
# end_row, end_col = min(img_h - 1, ymax + border_y), min(img_w - 1, xmax + border_x)
start_row, start_col = ymin - border_y, xmin - border_x
end_row, end_col = ymax + border_y, xmax + border_x
if start_row < 0 or start_col < 0 or end_row >= img_h or end_col >= img_w:
msg = 'Invalid border {} for box {} in image {} of size {}'.format(
[border_x, border_y], [xmin, ymin, xmax, ymax], filename,
[img_w, img_h])
if allow_skipping_images:
print('\n' + msg + '\n')
continue
else:
raise AssertionError(msg)
if mask:
n_masks += 1
w, h = end_col - start_col, end_row - start_row
patch_img = img[start_row:end_row, start_col:end_col, :]
if not disable_resizing:
if max_dim > 0:
if w > h:
out_w = max_dim
out_h = 0
else:
out_h = max_dim
out_w = 0
elif min_dim > 0:
if w < h:
out_w = min_dim
out_h = 0
else:
out_h = min_dim
out_w = 0
else:
out_w, out_h = out_mask_size
scale_x = float(out_w) / float(w)
scale_y = float(out_h) / float(h)
if scale_x == 0:
scale_x = scale_y
out_w = int(w * scale_x)
elif scale_y == 0:
scale_y = scale_x
out_h = int(h * scale_y)
try:
patch_img = cv2.resize(patch_img, (out_w, out_h))
# print('patch_img: {}'.format(patch_img.shape))
except cv2.error as e:
print('patch_img: {}'.format(patch_img.shape))
print('out_size: {}, {}'.format(start_row, start_col))
print('out_size: {}, {}'.format(end_row, end_col))
print('out_size: {}, {}'.format(out_w, out_h))
raise cv2.error(e)
else:
out_w, out_h = w, h
_label = label
if id_number is None:
id_number = -1
if id_number > 0:
_label = '{}_{}'.format(_label, id_number)
if enable_out_suffix:
out_fname = '{}_{}_{}'.format(filename_no_ext, obj_id, label)
else:
out_fname = filename_no_ext
_xmin, _ymin = int((xmin - start_col) * scale_x), int(
(ymin - start_row) * scale_y)
_xmax, _ymax = int((xmax - start_col) * scale_x), int(
(ymax - start_row) * scale_y)
if map_to_bbox:
if not img_written:
img_written = 1
out_img_path = os.path.join(img_root_path, filename)
cv2.imwrite(out_img_path, img)
if enable_xml_annotations:
imageShape = [xml_reader.height, xml_reader.width, 3]
xml_writer = PascalVocWriter(out_xml_path, filename,
imageShape)
if mask:
if enable_xml_annotations:
bndbox = [xmin, ymin, xmax, ymax]
xml_writer.addBndBox(bndbox[0], bndbox[1], bndbox[2],
bndbox[3], label, difficult,
bbox_source, id_number, score,
mask, mask_img)
raw_data = {
'target_id': int(id_number),
'filename': filename,
'width': img_w,
'height': img_h,
'class': label,
'xmin': xmin,
'ymin': ymin,
'xmax': xmax,
'ymax': ymax
}
if show_img:
cv2.rectangle(img, (xmin, ymin), (xmax, ymax), (0, 255, 0),
2)
disp_img = img
else:
img_out_fname = out_fname + '.jpg'
if mask:
out_img_path = os.path.join(img_root_path, img_out_fname)
cv2.imwrite(out_img_path, patch_img)
if enable_xml_annotations:
n_mask = len(mask)
_mask = []
for i in range(n_mask):
_mask.append([(mask[i][0] - start_col) * scale_x,
(mask[i][1] - start_row) * scale_y,
mask[i][2]])
imageShape = [xml_reader.height, xml_reader.width, 3]
xml_writer = PascalVocWriter(out_xml_path,
xml_reader.filename,
imageShape)
bndbox = [_xmin, _ymin, _xmax, _ymax]
xml_writer.addBndBox(_xmin, _ymin, _xmax, _ymax, label,
difficult, bbox_source, id_number,
score, _mask)
raw_data = {
'target_id': int(id_number),
'filename': img_out_fname,
'width': out_w,
'height': out_h,
'class': label,
'xmin': _xmin,
'ymin': _ymin,
'xmax': _xmax,
'ymax': _ymax
}
if show_img:
cv2.rectangle(patch_img, (_xmin, _ymin), (_xmax, _ymax),
(0, 255, 0), 2)
disp_img = patch_img
if mask:
if mask_img is None:
mask_img = np.zeros_like(img)
# print('border_x: {}'.format(border_x))
# print('border_y: {}'.format(border_y))
# print('scale_x: {}'.format(scale_x))
# print('scale_y: {}'.format(scale_y))
#
# print('xmin: {}'.format(xmin))
# print('ymin: {}'.format(ymin))
mask_pts = [[(x - xmin + border_x) * scale_x,
(y - ymin + border_y) * scale_y]
for x, y in mask_pts]
curr_mask = np.zeros_like(patch_img, dtype=np.uint8)
# print('mask_img: {}'.format(mask_img.shape))
mask_out_fname = out_fname + '.png'
# np.savetxt('mask_seq_mask_pts.txt', mask_pts, fmt='%.6f')
curr_mask = cv2.fillPoly(
curr_mask, np.array([
mask_pts,
], dtype=np.int32), mask_pix_val)
# print('min: {} max: {}'.format(
# np.min(mask_img.flatten()),
# np.max(mask_img.flatten()))
# )
if map_to_bbox:
mask_img = map_mask_to_bbox(
(xmin, ymin, xmax, ymax), curr_mask, fixed_ar,
out_border, mask_img.shape, mask_img)
else:
mask_img = curr_mask
out_mask_path = os.path.join(out_mask_root_path,
mask_out_fname)
cv2.imwrite(out_mask_path, mask_img)
_train_fnames.append((out_img_path, out_mask_path))
if show_img:
disp_mask_img = mask_img.copy()
if save_raw_mask:
disp_mask_img[disp_mask_img > 0] = 255
blended_img = np.asarray(
Image.blend(Image.fromarray(patch_img),
Image.fromarray(disp_mask_img), 0.5))
disp_img = np.concatenate(
(disp_img, disp_mask_img, blended_img), axis=1)
csv_raw.append(raw_data)
else:
test_csv_raw.append(raw_data)
if not map_to_bbox:
_test_fnames.append(out_img_path)
if show_img and not map_to_bbox:
# if _pause:
# print('frame {} :: {}'.format(file_id, filename))
cv2.imshow(win_name, disp_img)
k = cv2.waitKey(1 - _pause)
if k == ord('q'):
_exit = 1
break
elif k == 27:
_exit_seq = 1
break
elif k == 32:
_pause = 1 - _pause
obj_id += 1
if map_to_bbox and img is not None:
out_img_path = os.path.join(out_img_root_path, filename)
if save_train and mask_img is not None:
mask_out_fname = filename_no_ext + '.png'
out_mask_path = os.path.join(out_mask_root_path,
mask_out_fname)
cv2.imwrite(out_mask_path, mask_img)
if enable_xml_annotations:
out_xml_file = os.path.join(out_xml_path,
os.path.basename(file))
xml_writer.save(targetFile=out_xml_file)
_train_fnames.append((out_img_path, out_mask_path))
if show_img:
disp_mask_img = mask_img
if save_raw_mask:
disp_mask_img[disp_mask_img > 0] = 255
blended_img = np.asarray(
Image.blend(Image.fromarray(img),
Image.fromarray(disp_mask_img), 0.5))
disp_img = np.concatenate(
(disp_img, disp_mask_img, blended_img), axis=1)
elif save_test:
out_img_path = os.path.join(test_img_root_path, filename)
if out_img_path in _test_fnames:
raise IOError(
'Duplicate out_img_path: {}'.format(out_img_path))
_test_fnames.append(out_img_path)
if show_img and disp_img is not None:
cv2.imshow(win_name, disp_img)
k = cv2.waitKey(1 - _pause)
if k == ord('q'):
_exit = 1
break
elif k == 27:
break
elif k == 32:
_pause = 1 - _pause
if _exit:
break
sys.stdout.write(
'\rDone {:d}/{:d} files {:s} ({:d} masks found)'.format(
file_id + 1, n_files, filename, n_masks))
sys.stdout.flush()
if masks_per_seq > 0 and n_masks >= masks_per_seq:
break
sys.stdout.write('\n')
sys.stdout.flush()
if not _exit_seq and save_train and n_masks == 0:
raise IOError('\nNo masks found for {}\n'.format(seq_path))
train_csv_path = test_csv_path = ''
if csv_raw:
print('Saved {} labeled files in training sequence'.format(
len(csv_raw)))
train_csv_path = os.path.join(out_img_root_path, 'annotations.csv')
pd.DataFrame(csv_raw).to_csv(train_csv_path)
if test_csv_raw:
print('Saved {} unlabeled files in test sequence'.format(
len(test_csv_raw)))
test_csv_path = os.path.join(test_img_root_path, 'annotations.csv')
pd.DataFrame(test_csv_raw).to_csv(test_csv_path)
if show_img:
cv2.destroyWindow(win_name)
if save_train and train_fnames is not None:
train_fnames[
out_seq_name] = _train_fnames, train_root_path, csv_raw, train_csv_path
if save_test and test_fnames is not None:
test_fnames[
out_test_seq_name] = _test_fnames, test_img_root_path, test_csv_raw, test_csv_path
return n_masks
def saveBoxesTXT(_type, voc_path, class_dict, out_dir=''):
if _type == 0:
_type_str = 'mAP'
else:
_type_str = 'yolo'
if not voc_path or not os.path.isdir(voc_path):
print('Folder containing the loaded boxes does not exist')
return None
files = glob.glob(os.path.join(voc_path, '*.xml'))
n_files = len(files)
if n_files == 0:
print('No loaded boxes found')
return None
def convert_to_yolo(size, box):
dw = 1. / size[0]
dh = 1. / size[1]
x = (box[0] + box[1]) / 2.0
y = (box[2] + box[3]) / 2.0
w = box[1] - box[0]
h = box[3] - box[2]
x = x * dw
w = w * dw
y = y * dh
h = h * dh
return x, y, w, h
def getint(fn):
basename = os.path.basename(fn)
num = re.sub("\D", "", basename)
try:
return int(num)
except:
return 0
if len(files) > 0:
files = sorted(files, key=getint)
if not out_dir:
out_dir = os.path.join(os.path.dirname(voc_path), _type_str)
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
list_file = None
if _type == 1:
list_path = os.path.join(out_dir, 'list.txt')
list_file = open(list_path, 'w')
print('Loading VOC annotations from {:d} files at {:s}...'.format(n_files, voc_path))
print('Writing {} annotations to {:s}...'.format(_type_str, out_dir))
file_id = 0
n_boxes = 0
for file in files:
file_no_ext = os.path.splitext(os.path.basename(file))[0]
out_file_path = os.path.join(out_dir, '{}.txt'.format(file_no_ext))
out_file = open(out_file_path, 'w')
xml_reader = PascalVocReader(file)
shapes = xml_reader.getShapes()
img_width = xml_reader.width
img_height = xml_reader.height
for shape in shapes:
label, points, _, _, difficult, bbox_source, id_number, score, _, _ = shape
xmin, ymin = points[0]
xmax, ymax = points[2]
def clamp(x, min_value=0.0, max_value=1.0):
return max(min(x, max_value), min_value)
xmin = int(clamp(xmin, 0, img_width-1))
xmax = int(clamp(xmax, 0, img_width-1))
ymin = int(clamp(ymin, 0, img_height-1))
ymax = int(clamp(ymax, 0, img_height-1))
if _type == 0:
out_file.write('{:s} {:d} {:d} {:d} {:d}\n'.format(label, xmin, ymin, xmax, ymax))
else:
class_id = class_dict[label] + 1
bb = convert_to_yolo((xml_reader.width, xml_reader.height), [xmin, xmax, ymin, ymax])
out_file.write('{:d} {:f} {:f} {:f} {:f}\n'.format(class_id, bb[0], bb[1], bb[2], bb[3]))
if _type == 1:
list_file.write('{:s}\n'.format(xml_reader.filename))
n_boxes += 1
file_id += 1
sys.stdout.write('\rDone {:d}/{:d} files with {:d} boxes ({:d}x{:d})'.format(
file_id, n_files, n_boxes, img_width, img_height))
sys.stdout.flush()
out_file.close()
if _type == 1:
list_file.close()
sys.stdout.write('\n')
sys.stdout.flush()
return out_dir
def saveBoxesCSV(seq_path,
voc_path,
sources_to_include,
enable_mask,
img_ext='jpg',
verbose=True):
if not voc_path or not os.path.isdir(voc_path):
raise IOError(
'Folder containing the loaded boxes does not exist: {}'.format(
voc_path))
# return None
# src_files = [os.path.join(seq_path, k) for k in os.listdir(seq_path) if
# os.path.splitext(k.lower())[1][1:] == img_ext]
# src_files.sort(key=sortKey)
# seq_name = os.path.basename(img_path)
files = glob.glob(os.path.join(voc_path, '*.xml'))
n_files = len(files)
if n_files == 0:
print('No loaded boxes found')
return None
def getint(fn):
basename = os.path.basename(fn)
num = re.sub("\D", "", basename)
try:
return int(num)
except:
return 0
if len(files) > 0:
files = sorted(files, key=getint)
print('Loading annotations from {:d} files at {:s}...'.format(
n_files, voc_path))
file_id = 0
n_boxes = 0
csv_raw = []
sources_to_exclude = [
k[1:] for k in sources_to_include if k.startswith('!')
]
sources_to_include = [
k for k in sources_to_include if not k.startswith('!')
]
if sources_to_include:
print('Including only boxes from following sources: {}'.format(
sources_to_include))
if sources_to_exclude:
print('Excluding boxes from following sources: {}'.format(
sources_to_exclude))
for file in files:
xml_reader = PascalVocReader(file)
shapes = xml_reader.getShapes()
for shape in shapes:
label, points, _, _, difficult, bbox_source, id_number, score, mask_pts, _ = shape
if sources_to_include and bbox_source not in sources_to_include:
continue
if sources_to_exclude and bbox_source in sources_to_exclude:
continue
if id_number is None:
id_number = -1
xmin, ymin = points[0]
xmax, ymax = points[2]
filename = os.path.splitext(
os.path.basename(file))[0] + '.{}'.format(img_ext)
filename_from_xml = xml_reader.filename
# if filename != filename_from_xml:
# sys.stdout.write('\n\nImage file name from xml: {} does not match the expected one: {}'.format(
# filename_from_xml, filename))
# sys.stdout.flush()
raw_data = {
'target_id': int(id_number),
'filename': filename,
'width': xml_reader.width,
'height': xml_reader.height,
'class': label,
'xmin': int(xmin),
'ymin': int(ymin),
'xmax': int(xmax),
'ymax': int(ymax)
}
if enable_mask:
mask_txt = ''
if mask_pts is not None and mask_pts:
mask_txt = '{},{};'.format(*(mask_pts[0][:2]))
for _pt in mask_pts[1:]:
mask_txt = '{}{},{};'.format(mask_txt, _pt[0], _pt[1])
raw_data.update({'mask': mask_txt})
csv_raw.append(raw_data)
n_boxes += 1
file_id += 1
if verbose:
sys.stdout.write('\rDone {:d}/{:d} files with {:d} boxes'.format(
file_id, n_files, n_boxes))
sys.stdout.flush()
if verbose:
sys.stdout.write('\n')
sys.stdout.flush()
df = pd.DataFrame(csv_raw)
out_dir = os.path.dirname(voc_path)
out_file_path = os.path.join(out_dir, 'annotations.csv')
csv_columns = [
'target_id', 'filename', 'width', 'height', 'class', 'xmin', 'ymin',
'xmax', 'ymax'
]
if enable_mask:
csv_columns.append('mask')
df.to_csv(out_file_path, columns=csv_columns)
return out_file_path
def run(self):
path = self.msg['path']
frame_number = self.msg['frame_number']
width = self.msg['width']
height = self.msg['height']
channel = self.msg['channel']
bboxes = self.msg['bboxes']
scores = self.msg['scores']
labels = self.msg['labels']
bbox_source = self.msg['bbox_source']
id_numbers = self.msg['id_numbers']
last_frame_number = self.msg['last_frame_number']
trigger_tracking_request = self.msg['trigger_tracking_request']
num_frames = self.msg['num_frames']
new_box_from_pt = (bbox_source == 'single_object_tracker')
num_new_bboxes = len(bboxes)
if 'masks' in self.msg.keys():
masks = self.msg['masks']
else:
masks = None
trigger_batch_message = num_frames > 1
if not self.main_window.frames_reader or os.path.abspath(
path) != self.main_window.frames_reader.get_path():
print("Incorrect path!")
# send_self.msg_to_connection(dict(status="failure", error="Incorrect path!"), connection)
else:
if new_box_from_pt or self.save_boxes:
xml_path = self.main_window.get_annotation_path(frame_number)
if os.path.exists(xml_path):
try:
tVocParseReader = PascalVocReader(xml_path)
shapes = tVocParseReader.getShapes()
except:
shapes = []
else:
shapes = []
imageShape = [height, width, channel]
if isinstance(self.main_window.frames_reader, DirectoryReader):
img_full_path = self.main_window.mImgList[frame_number]
folder_name = os.path.dirname(img_full_path)
file_name = os.path.basename(img_full_path)
else:
folder_name = self.main_window.frames_reader.video_fn
file_name = str(frame_number)
tVocWriter = PascalVocWriter(folder_name, file_name,
imageShape) #
for label, points, _, _, difficult, _bbox_source, id_number, score, mask, mask_img in shapes:
existing_box_from_gt = _bbox_source == "ground_truth"
if new_box_from_pt and existing_box_from_gt and id_number == id_numbers[
0]:
print(
'Received duplicale target {:d} bbox for frame {:d} for which GT exists'
.format(id_numbers[0], frame_number))
return
bndbox = LabelFile.convertPoints2BndBox(points, label)
if existing_box_from_gt or _bbox_source != bbox_source or label == 'gate' or \
(_bbox_source == "single_object_tracker" and id_number != id_numbers[0]):
# Override previous bboxes from the same source
tVocWriter.addBndBox(bndbox[0], bndbox[1], bndbox[2],
bndbox[3], label, difficult,
_bbox_source, id_number, score,
mask, mask_img)
for j in range(num_new_bboxes):
bbox = bboxes[j]
xmin = bbox['xmin']
ymin = bbox['ymin']
xmax = bbox['xmax']
ymax = bbox['ymax']
label = labels[j]
score = scores[j]
difficulty = False
id_number = id_numbers[j]
if masks is not None:
# box_h = ymax - ymin
# box_w = xmax - xmin
# mask_img = np.array(masks[j]).reshape((box_h, box_w))
# mask_img = np.array(masks[j])
print('Loading mask from {:s}'.format(masks[j]))
# mask_img = cv2.imread(masks[j])
mask_img = np.load(masks[j])
if mask_img is None:
print('mask image could not be read')
else:
mask_img = None
tVocWriter.addBndBox(xmin, ymin, xmax, ymax, label,
difficulty, bbox_source, id_number,
score, None, mask_img)
# print('bbox: ', bbox)
tVocWriter.save(targetFile=xml_path)
if self.currIndex == frame_number:
self.reload_signal.emit()
if new_box_from_pt:
print('Received target {:d} bbox for frame {:d}'.format(
id_numbers[0], frame_number))
else:
self.update_status_signal.emit(
"Received {:d} {:s} bboxes for frame {:d}.".format(
num_new_bboxes, bbox_source, frame_number))
if last_frame_number == frame_number and not trigger_tracking_request:
self.update_status_signal.emit(
"Running finished for {:d} frames.".format(num_frames))
self.finish_signal.emit(trigger_batch_message)
