def openpose_convert_labelme_format(input_json,
input_image_path=None,
output_dir=None,
only_pose=False):
openPose_object = OpenPose_JSON(input_json)
labelme_object = Labelme_JSON()
image_name = utils_io_folder.get_file_name_from_path(
input_image_path) # like 00000001.jpg
labelme_object.imagePath = ""
# labelme_object.imageData = ""
labelme_object.imageData = utils_image.img_encode_to_base64(
input_image_path)
image_arr = utils_image.img_b64_to_arr(
utils_image.img_encode_to_base64(input_image_path))
labelme_object.imageHeight, labelme_object.imageWidth = image_arr.shape[
0], image_arr.shape[1]
for person in openPose_object.people:
track_id = person["person_id"]
pose_keypoints_2d = person["pose_keypoints_2d"]
hand_left_keypoints_2d = person["hand_left_keypoints_2d"]
hand_right_keypoints_2d = person["hand_right_keypoints_2d"]
if only_pose:
for index, keypoint in enumerate(pose_keypoints_2d):
labelme_object.shapes.append(
point_shape(track_id, pose_keypoints_order[index],
keypoint))
else:
for index, keypoint in enumerate(pose_keypoints_2d):
labelme_object.shapes.append(
point_shape(track_id, pose_keypoints_order[index],
keypoint))
for index, keypoint in enumerate(hand_left_keypoints_2d):
if index in hand_keep_point:
labelme_object.shapes.append(
point_shape(
track_id,
hand_keypoints_order[index].format("left_"),
keypoint))
for index, keypoint in enumerate(hand_right_keypoints_2d):
if index in hand_keep_point:
labelme_object.shapes.append(
point_shape(
track_id,
hand_keypoints_order[index].format("right_"),
keypoint))
if output_dir is None:
# 输出 默认在input_json的路径下创建labelme文件夹。
output_dir = os.path.join(
utils_io_folder.get_parent_folder_from_path(input_json))
# utils_io_folder.create_folder(output_dir)
output_path = os.path.join(
output_dir, utils_io_folder.replace_file_ext(image_name, "json"))
else:
utils_io_folder.create_folder(output_dir)
output_path = os.path.join(
output_dir, utils_io_folder.replace_file_ext(image_name, "json"))
utils_json.write_json_to_file(labelme_object.to_dict(), output_path)
def openSVAI_convert_labelme_format(input_json, input_images_dir=None, output_dir=None):
openSVAI_object = OpenSVAI_JSON(input_json)
for frame in openSVAI_object.frames:
labelme_object = Labelme_JSON()
image_info = frame["image"]
candidates_info = frame["candidates"]
image_name = image_info["name"] # like 00000001.jpg
if input_images_dir is None:
labelme_object.imagePath = os.path.join(image_info["folder"], image_info["name"])
else:
labelme_object.imagePath = os.path.join(input_images_dir, image_name)
labelme_object.imageData = utils_image.img_encode_to_base64(labelme_object.imagePath)
image_arr = utils_image.img_b64_to_arr(labelme_object.imageData)
labelme_object.imageHeight, labelme_object.imageWidth = image_arr.shape[0], image_arr.shape[1]
for candidate in candidates_info:
track_id = candidate['track_id'] # 对应 labelme的group_id
det_bbox = candidate["det_bbox"]
# det_bbox (open_SVAI) - rectangle (labelme)
labelme_object.shapes.append(rectangle_shape(track_id, det_bbox))
keypoints = numpy.array(candidate['pose_keypoints_2d']).reshape(-1, 3)
for index, keypoint in enumerate(keypoints):
# keypoint (open_SVAI) - point (labelme)
labelme_object.shapes.append(point_shape(track_id, keypoints_order[index], keypoint))
if output_dir is None:
# 输出 默认在input_json的路径下创建labelme文件夹。
output_dir = os.path.join(utils_io_folder.get_parent_folder_from_path(input_json), "labelme")
utils_io_folder.create_folder(output_dir)
output_path = os.path.join(output_dir, utils_io_folder.replace_file_ext(image_name, "json"))
else:
utils_io_folder.create_folder(output_dir)
output_path = os.path.join(output_dir, utils_io_folder.replace_file_ext(image_name, "json"))
utils_json.write_json_to_file(labelme_object.to_dict(), output_path)
def posetrack2017_convert_labelme_format(input_json, input_image_path_base,
output_dir_base):
"""
输入是一段视频的json 仅一个json文件，输出为每个帧的json文件
:param input_json:
:param input_image_path_base:
:param output_dir_base:
:return:
"""
annolist = utils_json.read_json_from_file(input_json)["annolist"]
for annotation in annolist:
# 处理一张图片
# openPose_object = OpenPose_JSON(input_json)
input_image_path = os.path.join(input_image_path_base,
annotation["image"][0]["name"])
labelme_object = Labelme_JSON()
image_name = utils_io_folder.get_file_name_from_path(
input_image_path) # like 00000001.jpg
labelme_object.imagePath = input_image_path
labelme_object.imageData = utils_image.img_encode_to_base64(
labelme_object.imagePath)
image_arr = utils_image.img_b64_to_arr(labelme_object.imageData)
labelme_object.imageHeight, labelme_object.imageWidth = image_arr.shape[
0], image_arr.shape[1]
people = annotation["annorect"]
for person in people:
track_id = person["track_id"]
annopoints = person["annopoints"][0]["point"]
for point in annopoints:
point_id = point["id"][0]
x = point["x"][0]
y = point["y"][0]
is_visible = point["is_visible"][0]
keypoint = [x, y]
labelme_object.shapes.append(
point_shape(
track_id,
joint_names[point_id] + "_vis_{}".format(is_visible),
keypoint))
for point_1 in annopoints:
point_1_id = point_1["id"][0]
x_1 = point_1["x"][0]
y_1 = point_1["y"][0]
for point_2 in annopoints:
point_2_id = point_2["id"][0]
x_2 = point_2["x"][0]
y_2 = point_2["y"][0]
for joint_pair in joint_pairs:
point_1_name = joint_names[point_1_id]
point_2_name = joint_names[point_2_id]
if joint_pair[0] == point_1_name and joint_pair[
1] == point_2_name:
labelme_object.shapes.append(
line_shape(track_id,
point_1_name + point_2_name,
[x_1, y_1], [x_2, y_2]))
path = annotation["image"][0]["name"]
output_dir = output_dir_base + path[path.index("/"):path.rindex("/"
)]
utils_io_folder.create_folder(output_dir)
output_path = os.path.join(
output_dir,
utils_io_folder.replace_file_ext(image_name, "json"))
utils_json.write_json_to_file(labelme_object.to_dict(),
output_path)