当前位置：首页 > news >正文

使用labelme进行实例分割标注

news 来源：原创 2025/9/12 3:20:01

前言

最近在学习实例分割算法，参考b站视频课教程，使用labelme标注数据集，在csdn找到相关教程进行数据集格式转换，按照相关目标检测网络对数据集格式的训练要求划分数据集。

1.使用labelme标注图片

在网上随便找了几张蘑菇图片，使用多边形进行标注，标签设置为mushroom,

注意：实例分割标注疑惑：一张图片同类物体不同实例，如图片中显示多个蘑菇，该标注为mushroom1, mushroom2,mushroom3;或者全都标注为mushroom

在查看了一些实例相关的文献，发现以上两种情况均存在

在使用maskrcnn的文献中，一些是同类物体不同实例：显示不同掩膜颜色；而另一些是同类物体不同实例：显示相同掩膜颜色。

最终决定选择全部标注为mushroom的原因：

1.参考github官方labelme的实例分割标注示例

E:\Github\github\labelme-main\examples\instance_segmentation\data_annotated

图片为2011_000006.jpg

labelme json标注信息,可以看到图片中有三个人，在json标签中三个实例同一种类（人）label均为person, 因此说明label指代是类别标签，而不是实例标签，如果标注为person1,perosn2,person3,则是说明有三种不同的类别，如果再标注另一张带有人的标签，也标注为perosn1,perosn2，则会出现每张图片不同的人，都有一个person1, 但是实际上他们类别都是person。因此这样标显然不合理。

{"version": "4.0.0","flags": {},"shapes": [{"label": "person","points": [[204.936170212766,108.56382978723406],[183.936170212766,141.56382978723406],[166.936170212766,150.56382978723406],[108.93617021276599,203.56382978723406],[92.93617021276599,228.56382978723406],[95.93617021276599,244.56382978723406],[105.93617021276599,244.56382978723406],[116.93617021276599,223.56382978723406],[163.936170212766,187.56382978723406],[147.936170212766,212.56382978723406],[117.93617021276599,222.56382978723406],[108.93617021276599,243.56382978723406],[100.93617021276599,325.56382978723406],[135.936170212766,329.56382978723406],[148.936170212766,319.56382978723406],[150.936170212766,295.56382978723406],[169.936170212766,272.56382978723406],[171.936170212766,249.56382978723406],[178.936170212766,246.56382978723406],[186.936170212766,225.56382978723406],[214.936170212766,219.56382978723406],[242.936170212766,157.56382978723406],[228.936170212766,146.56382978723406],[228.936170212766,125.56382978723406],[216.936170212766,112.56382978723406]],"group_id": null,"shape_type": "polygon","flags": {}},{"label": "person","points": [[271.936170212766,109.56382978723406],[249.936170212766,110.56382978723406],[244.936170212766,150.56382978723406],[215.936170212766,219.56382978723406],[208.936170212766,245.56382978723406],[214.936170212766,220.56382978723406],[188.936170212766,227.56382978723406],[170.936170212766,246.56382978723406],[170.936170212766,275.56382978723406],[221.936170212766,278.56382978723406],[233.936170212766,259.56382978723406],[246.936170212766,253.56382978723406],[245.936170212766,256.56382978723406],[242.936170212766,251.56382978723406],[262.936170212766,256.56382978723406],[304.936170212766,226.56382978723406],[297.936170212766,199.56382978723406],[308.936170212766,164.56382978723406],[296.936170212766,148.56382978723406]],"group_id": null,"shape_type": "polygon","flags": {}},{"label": "person","points": [[308.936170212766,115.56382978723406],[298.936170212766,145.56382978723406],[309.936170212766,166.56382978723406],[297.936170212766,200.56382978723406],[305.936170212766,228.56382978723406],[262.936170212766,258.56382978723406],[252.936170212766,284.56382978723406],[272.936170212766,291.56382978723406],[281.936170212766,250.56382978723406],[326.936170212766,235.56382978723406],[351.936170212766,239.56382978723406],[365.936170212766,223.56382978723406],[371.936170212766,187.56382978723406],[353.936170212766,168.56382978723406],[344.936170212766,143.56382978723406],[336.936170212766,115.56382978723406]],"group_id": null,"shape_type": "polygon","flags": {}},{"label": "chair","points": [[309.7054009819968,242.94844517184941],[282.7054009819968,251.94844517184941],[271.7054009819968,287.9484451718494],[175.70540098199677,275.9484451718494],[149.70540098199677,296.9484451718494],[151.70540098199677,319.9484451718494],[160.70540098199677,328.9484451718494],[165.54250204582655,375.38461538461536],[486.7054009819968,373.9484451718494],[498.7054009819968,336.9484451718494],[498.7054009819968,202.94844517184941],[454.7054009819968,193.94844517184941],[435.7054009819968,212.94844517184941],[368.7054009819968,224.94844517184941],[351.7054009819968,241.94844517184941]],"group_id": null,"shape_type": "polygon","flags": {}},{"label": "person","points": [[425.936170212766,82.56382978723406],[404.936170212766,109.56382978723406],[400.936170212766,114.56382978723406],[437.936170212766,114.56382978723406],[448.936170212766,102.56382978723406],[446.936170212766,91.56382978723406]],"group_id": null,"shape_type": "polygon","flags": {}},{"label": "__ignore__","points": [[457.936170212766,85.56382978723406],[439.936170212766,117.56382978723406],[477.936170212766,117.56382978723406],[474.936170212766,87.56382978723406]],"group_id": null,"shape_type": "polygon","flags": {}},{"label": "sofa","points": [[183.936170212766,140.56382978723406],[125.93617021276599,140.56382978723406],[110.93617021276599,187.56382978723406],[22.936170212765987,199.56382978723406],[18.936170212765987,218.56382978723406],[22.936170212765987,234.56382978723406],[93.93617021276599,239.56382978723406],[91.93617021276599,229.56382978723406],[110.93617021276599,203.56382978723406]],"group_id": 0,"shape_type": "polygon","flags": {}},{"label": "sofa","points": [[103.93617021276599,290.56382978723406],[58.93617021276599,303.56382978723406],[97.93617021276599,311.56382978723406]],"group_id": 0,"shape_type": "polygon","flags": {}},{"label": "sofa","points": [[348.936170212766,146.56382978723406],[472.936170212766,149.56382978723406],[477.936170212766,162.56382978723406],[471.936170212766,196.56382978723406],[453.936170212766,192.56382978723406],[434.936170212766,213.56382978723406],[368.936170212766,226.56382978723406],[375.936170212766,187.56382978723406],[353.936170212766,164.56382978723406]],"group_id": 0,"shape_type": "polygon","flags": {}},{"label": "sofa","points": [[246.936170212766,252.56382978723406],[219.936170212766,277.56382978723406],[254.936170212766,287.56382978723406],[261.936170212766,256.56382978723406]],"group_id": 0,"shape_type": "polygon","flags": {}}],"imagePath": "2011_000006.jpg","imageData": null,"imageHeight": 375,"imageWidth": 500
}

在github/labelme的官网中，作者是按上图进行标注的，三个person

2.github官方标注实例中voc数据集和coco数据集

labelme/voc数据集格式

在github\labelme-main\examples\instance_segmentation\data_dataset_voc\SegmentationObject

文件中可以看到，实例分割标签中，三个人掩膜分别是不同的颜色

文件github\labelme-main\examples\instance_segmentation\data_dataset_voc\SegmentationObjectVisualization

在可视化实例分割标签中，可以看到不同人对应不同的数字和颜色

（红色1，绿色2，黄色3）

labelme/coco数据集格式

在github\labelme-main\examples\instance_segmentation\data_dataset_coco\Visualization文件夹中可以看到三个人，掩膜显示不同的颜色，标签均显示为person

查看其annotations.json 标签文件，发现label 也都是person, 那怎么区分不同实例呢？

发现它这里是用id 来区分不同实例，每个id 代表不同种类不同实例的一个独立物体编号（不考虑背景）

观察该图片，结合 annotations.json 标签文件分析

图片及其	image_id	物体	annotations.json 中实例对应id
2011_000003.jpg	0	(category_id=15) person	0
		person	1
		(category_id=5) bottel	2
2011_000025.jpg	1	(category_id=6) bus	3
		bus	4
		(category_id=7) car	5
2011_000006.jpg	2	person	6
		person	7
		person	8
		(category_id=9) chair	9
		person	10
		(category_id=18) sofa	11

注意：group_id

labelme标注生成的json标签文件中，可以看到4个 label为sofa，并且group_id 均为0，这表示这4个标注部分均属于同一个实例，而不是四个相互独立的实例（因为图片显示只有一张沙发，但是由于人遮挡关系，被分割为4个部分）

在voc数据集格式实例分割标签.png中，可以看到sofa的4个部分是同一个颜色（灰色），均属于图片中第7个实例（背景是第0个实例）

在coco数据集格式annotations.json文件中可以看到，关于sofa的标注信息，内部使用四个分割列表组成，说明该实例对象有被遮挡，而成为几个不连贯的部分；而其它实例对象是由一个分割列表组成，说明该实例对象没有被遮挡，内部连贯。

{"id": 11, "image_id": 2, "category_id": 18, "segmentation": 
[[183.936170212766, 140.56382978723406, 125.93617021276599, 140.56382978723406, 110.93617021276599, 187.56382978723406, 22.936170212765987, 199.56382978723406, 18.936170212765987, 218.56382978723406, 22.936170212765987, 234.56382978723406, 93.93617021276599, 239.56382978723406, 91.93617021276599, 229.56382978723406, 110.93617021276599, 203.56382978723406], [103.93617021276599, 290.56382978723406, 58.93617021276599, 303.56382978723406, 97.93617021276599, 311.56382978723406],[348.936170212766, 146.56382978723406, 472.936170212766, 149.56382978723406, 477.936170212766, 162.56382978723406, 471.936170212766, 196.56382978723406, 453.936170212766, 192.56382978723406, 434.936170212766, 213.56382978723406, 368.936170212766, 226.56382978723406, 375.936170212766, 187.56382978723406, 353.936170212766, 164.56382978723406], [246.936170212766, 252.56382978723406, 219.936170212766, 277.56382978723406, 254.936170212766, 287.56382978723406, 261.936170212766, 256.56382978723406]],"area": 14001.0, "bbox": [18.0, 140.0, 460.0, 172.0],"iscrowd": 0}

使用labelme自带的格式转换代码进行格式转换

在github\labelme-main\examples\instance_segmentation 文件夹下，

有两个py文件

labelme2coco.py	将labelme标注的json标签转换为coco数据集格式
labelme2voc.py	将labelme标注的json标签转换为voc数据集格式

labelme2voc.py使用

在使用labelme-多边形标注完图片后，我们将得到图片jpg，以及标签json （存放于data_annotated文件夹）

在转化前，需要再建立一个labels.txt文件，存放类别

注意：如果标注多个类别，该如何排序，观察labelme给示例，发现类别的排序与 coco数据集格式的 supercategory 的ID 序号对应。

labels.txt内容如下，_background_ 对应序号supercategory_id 为0， person 对应supercategory_id 为15 。

__ignore__
_background_
aeroplane
bicycle
bird
boat
bottle
bus
car
cat
chair
cow
diningtable
dog
horse
motorbike
person
potted plant
sheep
sofa
train
tv/monitor

coco数据集 annotations.json文件内容

"categories": [{"supercategory": null, "id": 0, "name": "_background_"}, {"supercategory": null, "id": 1, "name": "aeroplane"}, {"supercategory": null, "id": 2, "name": "bicycle"}, {"supercategory": null, "id": 3, "name": "bird"}, {"supercategory": null, "id": 4, "name": "boat"}, {"supercategory": null, "id": 5, "name": "bottle"}, {"supercategory": null, "id": 6, "name": "bus"}, {"supercategory": null, "id": 7, "name": "car"}, {"supercategory": null, "id": 8, "name": "cat"}, {"supercategory": null, "id": 9, "name": "chair"}, {"supercategory": null, "id": 10, "name": "cow"}, {"supercategory": null, "id": 11, "name": "diningtable"}, {"supercategory": null, "id": 12, "name": "dog"}, {"supercategory": null, "id": 13, "name": "horse"}, {"supercategory": null, "id": 14, "name": "motorbike"}, {"supercategory": null, "id": 15, "name": "person"}, {"supercategory": null, "id": 16, "name": "potted plant"}, {"supercategory": null, "id": 17, "name": "sheep"}, {"supercategory": null, "id": 18, "name": "sofa"}, {"supercategory": null, "id": 19, "name": "train"}, {"supercategory": null, "id": 20, "name": "tv/monitor"}]}

因此，创建labels.txt文件内容如下：

将下面四个文件存放在同一个文件夹，程序py文件从github\labelme-main\examples\instance_segmentation 复制

上面四个文件存放于文件夹test，在该文件夹下打开终端，激活虚拟环境，输入指令

python labelme2voc.py data_annotated data_dataset_voc --labels labels.txt

即可得到 VOC数据集格式

labelme2voc.py代码

#!/usr/bin/env pythonfrom __future__ import print_functionimport argparse
import glob
import os
import os.path as osp
import sysimport imgviz
import numpy as npimport labelmedef main():parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)parser.add_argument("input_dir", help="Input annotated directory")parser.add_argument("output_dir", help="Output dataset directory")parser.add_argument("--labels", help="Labels file or comma separated text", required=True)parser.add_argument("--noobject", help="Flag not to generate object label", action="store_true")parser.add_argument("--nonpy", help="Flag not to generate .npy files", action="store_true")parser.add_argument("--noviz", help="Flag to disable visualization", action="store_true")args = parser.parse_args()if osp.exists(args.output_dir):print("Output directory already exists:", args.output_dir)sys.exit(1)os.makedirs(args.output_dir)os.makedirs(osp.join(args.output_dir, "JPEGImages"))os.makedirs(osp.join(args.output_dir, "SegmentationClass"))if not args.nonpy:os.makedirs(osp.join(args.output_dir, "SegmentationClassNpy"))if not args.noviz:os.makedirs(osp.join(args.output_dir, "SegmentationClassVisualization"))if not args.noobject:os.makedirs(osp.join(args.output_dir, "SegmentationObject"))if not args.nonpy:os.makedirs(osp.join(args.output_dir, "SegmentationObjectNpy"))if not args.noviz:os.makedirs(osp.join(args.output_dir, "SegmentationObjectVisualization"))print("Creating dataset:", args.output_dir)if osp.exists(args.labels):with open(args.labels) as f:labels = [label.strip() for label in f if label]else:labels = [label.strip() for label in args.labels.split(",")]class_names = []class_name_to_id = {}for i, label in enumerate(labels):class_id = i - 1  # starts with -1class_name = label.strip()class_name_to_id[class_name] = class_idif class_id == -1:assert class_name == "__ignore__"continueelif class_id == 0:assert class_name == "_background_"class_names.append(class_name)class_names = tuple(class_names)print("class_names:", class_names)out_class_names_file = osp.join(args.output_dir, "class_names.txt")with open(out_class_names_file, "w") as f:f.writelines("\n".join(class_names))print("Saved class_names:", out_class_names_file)for filename in sorted(glob.glob(osp.join(args.input_dir, "*.json"))):print("Generating dataset from:", filename)label_file = labelme.LabelFile(filename=filename)base = osp.splitext(osp.basename(filename))[0]out_img_file = osp.join(args.output_dir, "JPEGImages", base + ".jpg")out_clsp_file = osp.join(args.output_dir, "SegmentationClass", base + ".png")if not args.nonpy:out_cls_file = osp.join(args.output_dir, "SegmentationClassNpy", base + ".npy")if not args.noviz:out_clsv_file = osp.join(args.output_dir,"SegmentationClassVisualization",base + ".jpg",)if not args.noobject:out_insp_file = osp.join(args.output_dir, "SegmentationObject", base + ".png")if not args.nonpy:out_ins_file = osp.join(args.output_dir, "SegmentationObjectNpy", base + ".npy")if not args.noviz:out_insv_file = osp.join(args.output_dir,"SegmentationObjectVisualization",base + ".jpg",)img = labelme.utils.img_data_to_arr(label_file.imageData)imgviz.io.imsave(out_img_file, img)cls, ins = labelme.utils.shapes_to_label(img_shape=img.shape,shapes=label_file.shapes,label_name_to_value=class_name_to_id,)ins[cls == -1] = 0  # ignore it.# class labellabelme.utils.lblsave(out_clsp_file, cls)if not args.nonpy:np.save(out_cls_file, cls)if not args.noviz:clsv = imgviz.label2rgb(cls,imgviz.rgb2gray(img),label_names=class_names,font_size=15,loc="rb",)imgviz.io.imsave(out_clsv_file, clsv)if not args.noobject:# instance labellabelme.utils.lblsave(out_insp_file, ins)if not args.nonpy:np.save(out_ins_file, ins)if not args.noviz:instance_ids = np.unique(ins)instance_names = [str(i) for i in range(max(instance_ids) + 1)]insv = imgviz.label2rgb(ins,imgviz.rgb2gray(img),label_names=instance_names,font_size=15,loc="rb",)imgviz.io.imsave(out_insv_file, insv)if __name__ == "__main__":main()

更改版，如果不想在终端使用命令行的形式运行，也可以在程序中输入文件夹路径直接运行

# 2025.4.12
"""注意事项：
确保 labels.txt 文件存在且格式正确输出目录不能预先存在（脚本会自动创建）路径中的反斜杠建议使用原始字符串（字符串前加r）如果出现权限问题，请以管理员身份运行
"""# !/usr/bin/env pythonfrom __future__ import print_functionimport glob
import os
import os.path as osp
import sysimport imgviz
import numpy as npimport labelmedef main():# ====================== 需要修改的配置 ======================# 输入标注目录（包含.json文件）input_dir = r"你的标注文件目录（包含.json文件）"# 输出目录（会自动创建）output_dir = r"输出目录（会自动创建）"# 标签文件路径labels_path = r"标签文件路径（例如：labels.txt）"# 功能开关（True表示启用，False表示禁用）noobject = False  # 是否不生成实例分割nonpy = False  # 是否不生成.npy文件noviz = False  # 是否不生成可视化结果# noobject = True  # 设置为True则不生成实例分割相关文件# nonpy = True     # 设置为True则不生成.npy文件# noviz = True     # 设置为True则不生成可视化图片# ===========================================================# 检查输出目录if osp.exists(output_dir):print("Output directory already exists:", output_dir)sys.exit(1)os.makedirs(output_dir)# 创建子目录os.makedirs(osp.join(output_dir, "JPEGImages"))os.makedirs(osp.join(output_dir, "SegmentationClass"))if not nonpy:os.makedirs(osp.join(output_dir, "SegmentationClassNpy"))if not noviz:os.makedirs(osp.join(output_dir, "SegmentationClassVisualization"))if not noobject:os.makedirs(osp.join(output_dir, "SegmentationObject"))if not nonpy:os.makedirs(osp.join(output_dir, "SegmentationObjectNpy"))if not noviz:os.makedirs(osp.join(output_dir, "SegmentationObjectVisualization"))print("Creating dataset:", output_dir)# 读取标签with open(labels_path) as f:labels = [label.strip() for label in f if label]# 处理类别映射class_names = []class_name_to_id = {}for i, label in enumerate(labels):class_id = i - 1  # starts with -1class_name = label.strip()class_name_to_id[class_name] = class_idif class_id == -1:assert class_name == "__ignore__"continueelif class_id == 0:assert class_name == "_background_"class_names.append(class_name)class_names = tuple(class_names)print("class_names:", class_names)# 保存类别名称out_class_names_file = osp.join(output_dir, "class_names.txt")with open(out_class_names_file, "w") as f:f.write("\n".join(class_names))print("Saved class_names:", out_class_names_file)# 处理每个标注文件for filename in sorted(glob.glob(osp.join(input_dir, "*.json"))):print("Generating dataset from:", filename)label_file = labelme.LabelFile(filename=filename)base = osp.splitext(osp.basename(filename))[0]out_img_file = osp.join(output_dir, "JPEGImages", base + ".jpg")out_clsp_file = osp.join(output_dir, "SegmentationClass", base + ".png")# 生成类别标签img = labelme.utils.img_data_to_arr(label_file.imageData)imgviz.io.imsave(out_img_file, img)cls, ins = labelme.utils.shapes_to_label(img_shape=img.shape,shapes=label_file.shapes,label_name_to_value=class_name_to_id,)ins[cls == -1] = 0  # 忽略被标记的区域# 保存类别分割结果labelme.utils.lblsave(out_clsp_file, cls)if not nonpy:np.save(osp.join(output_dir, "SegmentationClassNpy", base + ".npy"), cls)if not noviz:clsv = imgviz.label2rgb(cls,imgviz.rgb2gray(img),label_names=class_names,font_size=15,loc="rb",)imgviz.io.imsave(osp.join(output_dir, "SegmentationClassVisualization", base + ".jpg"),clsv,)# 生成实例分割结果if not noobject:out_insp_file = osp.join(output_dir, "SegmentationObject", base + ".png")labelme.utils.lblsave(out_insp_file, ins)if not nonpy:np.save(osp.join(output_dir, "SegmentationObjectNpy", base + ".npy"), ins)if not noviz:instance_ids = np.unique(ins)instance_names = [str(i) for i in range(max(instance_ids) + 1)]insv = imgviz.label2rgb(ins,imgviz.rgb2gray(img),label_names=instance_names,font_size=15,loc="rb",)imgviz.io.imsave(osp.join(output_dir, "SegmentationObjectVisualization", base + ".jpg"),insv,)if __name__ == "__main__":main()

data_dataset_voc文件夹内容如下：

data_dataset_voc\JPEGImages（存放jpg图片）

data_dataset_voc\SegmentationClass （存放语义分割标签）

data_dataset_voc\SegmentationClassNpy （存放语义分割标签的npy格式文件，好像训练时没有用到）

data_dataset_voc\SegmentationClassVisualization （语义分割标签-可视化）

data_dataset_voc\SegmentationObject （实例分割标签）

data_dataset_voc\SegmentationObjectNpy （存放实例分割标签的npy格式文件）

data_dataset_voc\SegmentationObjectVisualization （实例分割标签-可视化）

class_names.txt

内容

至此，labelme2voc.py 内容结束，但是在使用相关实例分割网络训练（mask-rcnn）的过程中，发现还需要xml文件，需要将labelme_json标签转换为xml标签，这个后续在其他文章再说。

labelme2coco.py使用

与上述一样，

在数据集文件夹中打开终端输入命令行参数

python  labelme2coco.py data_annotated data_dataset_coco --labels labels.txt

或者像上述更改版代码那样，进行修改，直接输入路径值

data_dataset_coco文件夹内容如下：


data_dataset_coco|__|——JPEGImages        图片.jpg|——Visualization     可视化实例分割标签 .png|——annotations.json  json标签信息 (所有图片)

data_dataset_coco\JPEGImages

data_dataset_coco\Visualization

data_dataset_coco\annotations.json

内容太长，就不展示了

参考资料：

1.github-labelme

2.labelme使用指南（转VOC、coco数据集）_哔哩哔哩_bilibili

3.deepseek.ai

前言