labelme进行关键点标注并转换为yolo格式

1、labelme安装和打开

在python3.9及以上环境中安装labelme，labelme要用到pyqt，所以在使用labelme之前要安装pyqt。

pip install pyqt
pip install labelme

进入存放标注图像文件夹的目录，提前设置要标注的文件夹、具体标签以及存放的文件夹。其中image是图像文件夹，labels是要存放标签的文件夹，labels.txt是提前设置的标签。

labelme image --output labels --labels labels.txt

打开的界面如下图所示

2、边界框和关键点标注

在标注时，要将目标框和对应的关键点分为一组，关键点也类似


标注完后记得点击上面save按钮

3、将lamelme的json格式转成yolo可以使用的txt格式

转换代码如下，可以将json格式根据group_id将目标和关键点对应起来，然后转成txt格式，每行格式为<class_id> <x_center> <y_center> <kp1_x> <kp1_y> <kp1_vis> <kp2_x> <kp2_y> <kp2_vis> … <kpN_x> <kpN_y> <kpN_vis>

import os
import json
from PIL import Imagedef convert_labelme_to_yolo_keypoints(json_path, image_dir, output_dir, class_id=0, num_keypoints=7):os.makedirs(output_dir, exist_ok=True)with open(json_path, 'r') as f:data = json.load(f)image_file = data.get("imagePath", os.path.basename(json_path).replace(".json", ".jpg"))image_path = os.path.join(image_dir, image_file)image = Image.open(image_path)width, height = image.sizeshapes = data['shapes']objects = {}  # key: group_id, value: {'bbox': [...], 'keypoints': {}}for shape in shapes:g_id = shape.get("group_id")if g_id is None:continuelabel = shape['label']shape_type = shape['shape_type']pts = shape['points']if g_id not in objects:objects[g_id] = {'bbox': None, 'keypoints': {}}if shape_type == "rectangle" and label == "bolt":x1, y1 = pts[0]x2, y2 = pts[1]objects[g_id]['bbox'] = [min(x1, x2), min(y1, y2), max(x1, x2), max(y1, y2)]elif shape_type == "point":try:kp_id = int(label)objects[g_id]['keypoints'][kp_id] = pts[0]except ValueError:pass  # skip if label is not a number# Write YOLO-style .txttxt_name = os.path.basename(json_path).replace('.json', '.txt')out_path = os.path.join(output_dir, txt_name)with open(out_path, 'w') as f:for obj in objects.values():if obj['bbox'] is None:continue  # skip if no boxx1, y1, x2, y2 = obj['bbox']x_center = (x1 + x2) / 2 / widthy_center = (y1 + y2) / 2 / heightbox_w = (x2 - x1) / widthbox_h = (y2 - y1) / heightkeypoints = []for i in range(1, num_keypoints + 1):if i in obj['keypoints']:x, y = obj['keypoints'][i]keypoints += [x / width, y / height, 2]else:keypoints += [0.0, 0.0, 0]  # 不存在该关键点line = f"{class_id} {x_center:.6f} {y_center:.6f} {box_w:.6f} {box_h:.6f} " + \" ".join([f"{kp:.6f}" if isinstance(kp, float) else str(kp) for kp in keypoints])f.write(line + '\n')print(f"✅ 转换完成: {out_path}")json_dir = r"path\to\labels"
image_dir = r"path\to\images"
output_dir = r"path\to\labels"for file in os.listdir(json_dir):if file.endswith(".json"):json_path = os.path.join(json_dir, file)convert_labelme_to_yolo_keypoints(json_path, image_dir, output_dir)

4、将数据和标签按照9比1分为训练集和测试集

代码如下

import os
import random
import shutil# 原始文件夹路径
images_dir = 'path\to\images'
labels_dir = 'path\to\labels'# 目标文件夹路径
train_image_dir = r'path\to\images\train'
val_image_dir = r'path\to\images\val'
train_label_dir = r'path\to\labels\train'
val_label_dir = r'path\to\labels\val'# 创建输出文件夹
os.makedirs(train_image_dir, exist_ok=True)
os.makedirs(val_image_dir, exist_ok=True)
os.makedirs(train_label_dir, exist_ok=True)
os.makedirs(val_label_dir, exist_ok=True)# 获取图像文件列表
image_files = [f for f in os.listdir(images_dir) if f.lower().endswith(('.jpg', '.jpeg', '.png'))]# 打乱并划分
random.shuffle(image_files)
split_idx = int(0.9 * len(image_files))
train_files = image_files[:split_idx]
val_files = image_files[split_idx:]# 拷贝函数
def copy_split(file_list, img_src, lbl_src, img_dst, lbl_dst):for fname in file_list:# 拷贝图像shutil.copy(os.path.join(img_src, fname), os.path.join(img_dst, fname))# 拷贝标签label_name = os.path.splitext(fname)[0] + '.txt'label_src_path = os.path.join(lbl_src, label_name)label_dst_path = os.path.join(lbl_dst, label_name)if os.path.exists(label_src_path):shutil.copy(label_src_path, label_dst_path)else:print(f"⚠️ 缺少标签: {label_src_path}")# 执行拷贝
copy_split(train_files, images_dir, labels_dir, train_image_dir, train_label_dir)
copy_split(val_files, images_dir, labels_dir, val_image_dir, val_label_dir)print("✅ 划分完成！图像和标签已按 9:1 存放到 images/train, images/val 和 labels/train, labels/val。")