当使用自建数据库对YOLO等模型进行训练时有时候会碰到无法直接使用.json文件进行训练的问题而labelme有时候标出来是.json格式。这里提供一个工具脚本在训练前先运行一遍可以有效解决该问题。该脚本原本用于YOLO-v8-seg模型训练这边类别数和名称清根据自己需要进行设置。该脚本只依赖标准库不用pip一堆适用范围说明支持所有有实例分割功能的模型即名称带-seg的版本但不支持没有的import os import json import random def convert_labelme_to_yolo_seg(json_dir, img_dir, output_dir, train_ratio0.8, keep_classesNone): if keep_classes is None: keep_classes [您的分类类别1, 您的分类类别2, 您的分类类别3, 您的分类类别4] class_to_id {name: idx for idx, name in enumerate(keep_classes)} num_classes len(class_to_id) # 创建输出子目录 for sub in [images/train, images/val, labels/train, labels/val]: os.makedirs(os.path.join(output_dir, sub), exist_okTrue) # 收集所有 JSON 文件 json_files [f for f in os.listdir(json_dir) if f.endswith(.json)] if not json_files: print(错误没有找到 JSON 文件) return random.shuffle(json_files) split_idx int(len(json_files) * train_ratio) train_files json_files[:split_idx] val_files json_files[split_idx:] def process_files(file_list, subset): for json_file in file_list: json_path os.path.join(json_dir, json_file) with open(json_path, r, encodingutf-8) as f: data json.load(f) # 获取图像文件名 img_name data.get(imagePath) if not img_name: base json_file.replace(.json, ) if os.path.exists(os.path.join(img_dir, base .jpg)): img_name base .jpg elif os.path.exists(os.path.join(img_dir, base .png)): img_name base .png else: print(f跳过 {json_file}找不到图像) continue # 复制图像手动读写避免 shutil src_img os.path.join(img_dir, img_name) dst_img os.path.join(output_dir, images, subset, img_name) if os.path.exists(src_img): with open(src_img, rb) as fsrc: with open(dst_img, wb) as fdst: fdst.write(fsrc.read()) else: print(f图像不存在{src_img}跳过) continue # 生成标签文件 label_file json_file.replace(.json, .txt) label_path os.path.join(output_dir, labels, subset, label_file) img_w data.get(imageWidth) img_h data.get(imageHeight) if img_w is None or img_h is None: print(f跳过 {json_file}缺少图像尺寸) continue with open(label_path, w) as f: for shape in data.get(shapes, []): label shape.get(label) if label not in class_to_id: continue class_id class_to_id[label] shape_type shape.get(shape_type, polygon) points shape.get(points, []) # 矩形 → 四边形四个顶点 if shape_type rectangle and len(points) 2: x1, y1 points[0] x2, y2 points[1] points [[x1, y1], [x2, y1], [x2, y2], [x1, y2]] if len(points) 3: print(f警告{json_file} 中 {label} 点数不足3跳过) continue # 归一化坐标并写入 norm [] for (x, y) in points: nx max(0.0, min(1.0, x / img_w)) ny max(0.0, min(1.0, y / img_h)) norm.extend([nx, ny]) f.write(f{class_id} .join(f{p:.6f} for p in norm) \n) # 若标签文件为空则删除 if os.path.getsize(label_path) 0: os.remove(label_path) print(f注意{json_file} 没有有效标注已删除空标签文件) print(f训练集: {len(train_files)} 张) process_files(train_files, train) print(f验证集: {len(val_files)} 张) process_files(val_files, val) # 生成 dataset.yaml yaml_path os.path.join(output_dir, dataset.yaml) with open(yaml_path, w) as f: f.write(fpath: .\n) f.write(ftrain: images/train\n) f.write(fval: images/val\n) f.write(fnc: {num_classes}\n) f.write(fnames: {keep_classes}\n) print(f已生成 {yaml_path}) if __name__ __main__: convert_labelme_to_yolo_seg( json_dirdata/training_data/annotations, img_dirdata/training_data/image, output_dirdata/yolo_dataset, train_ratio0.8, keep_classes[您的分类类别1, 您的分类类别2, 您的分类类别3, 您的分类类别4] )