【语义分割】用labelme制作VOC数据集

https://blog.csdn.net/Maisie_Nan/article/details/101026814

使用pip安装labelme；
使用cmd启动labelme标注界面；
labelme
labels.txt存放分类数据。

自己修改下labelme2voc.py

python labelme2voc.py --labels=labels.txt input_folder data_dataset_voc

    # -*- coding: utf-8 -*-
    #!/usr/bin/env python
    """
    Created on Thu Sep 19 15:46:16 2019
    @author: Andrea
    """
     
     
     
    from __future__ import print_function
     
    import argparse
    import glob
    import json
    import os
    import os.path as osp
    import sys
     
    import numpy as np
    import PIL.Image
     
    import labelme
    from sklearn.model_selection import train_test_split
     
     
    def 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', required=True)
        args = parser.parse_args()
     
     
        if osp.exists(args.output_dir):
            print('Output directory already exists:', args.output_dir)
    #        sys.exit(1)
        else:
            os.makedirs(args.output_dir)
            os.makedirs(osp.join(args.output_dir, 'JPEGImages'))
            os.makedirs(osp.join(args.output_dir, 'SegmentationClass'))
        #    os.makedirs(osp.join(args.output_dir, 'SegmentationClassPNG'))
            os.makedirs(osp.join(args.output_dir, 'SegmentationClassVisualization'))
        saved_path = args.output_dir
        if not os.path.exists(os.path.join(saved_path , 'ImageSets','Segmentation')):
            os.makedirs(os.path.join(saved_path , 'ImageSets','Segmentation'))
        print('Creating dataset:', args.output_dir)
     
        class_names = []
        class_name_to_id = {}
        for i, line in enumerate(open(args.labels).readlines()):
            print(i)
            class_id = i+1 # starts with -1
            class_name = line.strip()
            class_name_to_id[class_name] = class_id
            if class_id == -1:
                assert class_name == '__ignore__'
                continue
            elif 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)
     
        colormap = labelme.utils.label_colormap(255)
     
        for label_file in glob.glob(osp.join(args.input_dir, '*.json')):
            print('Generating dataset from:', label_file)
            try:
                with open(label_file) as f:
                    base = osp.splitext(osp.basename(label_file))[0]
                    out_img_file = osp.join(
                        args.output_dir, 'JPEGImages', base + '.jpg')
        #            out_lbl_file = osp.join(
        #                args.output_dir, 'SegmentationClass', base + '.npy')
        #                args.output_dir, 'SegmentationClass', base + '.npy')
                    out_png_file = osp.join(
                        args.output_dir, 'SegmentationClass', base + '.png')
                    out_viz_file = osp.join(
                        args.output_dir,
                        'SegmentationClassVisualization',
                        base + '.jpg',
                    )
        
                    data = json.load(f)
        
                    img_file = osp.join(label_file.split('.json')[0]+'.jpg')
                    img = np.asarray(PIL.Image.open(img_file))
                    PIL.Image.fromarray(img).save(out_img_file)
        
        
                    print('class_name_to_id:',class_name_to_id)
                    lbl = labelme.utils.shapes_to_label(
                        img_shape=img.shape,
                        shapes=data['shapes'],
                        label_name_to_value=class_name_to_id,
                    )
                    labelme.utils.lblsave(out_png_file, lbl)
        
        #            np.save(out_lbl_file, lbl)
        
                    viz = labelme.utils.draw_label(
                        lbl, img, class_names, colormap=colormap)
                    PIL.Image.fromarray(viz).save(out_viz_file)
            except:
                with open('wrongdata.txt','w') as f:
                    f.write(label_file+'\n')
                print('这张图像有错误')
                continue
                
        #6.split files for txt
        txtsavepath = os.path.join(saved_path , 'ImageSets','Segmentation')
        ftrainval = open(os.path.join(txtsavepath,'trainval.txt'), 'w')
        ftest = open(os.path.join(txtsavepath,'test.txt'), 'w')
        ftrain = open(os.path.join(txtsavepath,'train.txt'), 'w')
        fval = open(os.path.join(txtsavepath,'val.txt'), 'w')
        total_files = os.listdir(osp.join(args.output_dir, 'SegmentationClass'))
        total_files = [i.split("/")[-1].split(".png")[0] for i in total_files]
        #test_filepath = ""
        for file in total_files:
            ftrainval.write(file + "\n")
        #test
        #for file in os.listdir(test_filepath):
        #    ftest.write(file.split(".jpg")[0] + "\n")
        #split
        train_files,val_files = train_test_split(total_files,test_size=0.15,random_state=42)
        #train
        for file in train_files:
            ftrain.write(file + "\n")
        #val
        for file in val_files:
            fval.write(file + "\n")
        
        ftrainval.close()
        ftrain.close()
        fval.close()
        ftest.close()
     
     
     
    if __name__ == '__main__':
        main()

运行完毕生成VOC数据集。

如果需要制作【目标检测】数据集：labelme 标注矩形检测数据格式转 VOC 数据集格式。

可参考：http://spytensor.com/index.php/archives/35/?hutoto=th9jn1&labule=dmrxn1
————————————————
版权声明：本文为CSDN博主「Mein_Augenstern」的原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/Maisie_Nan/article/details/101026814