离线半自动标注流程

技术文档

前言

• 为方便标注，加快标注进度，本教程使用CVAT的标签导入功能，进行离线的半自动标注
• 本教程基于Darknet版的Yolov4预训练模型实现半自动标签生成
• 本教程所有路径均在10.0.10.56服务器上。

流程说明

1. 用Yolov4在COCO上的预训练模型对测试集生成json格式的标签文件
2. 用/ssd01/wanghuijiao/dataset/Summary/yolor_json2yolo.py脚本将json文件转换为yolo格式，此脚本可以实现在原有的yolo标签文件内容后进行添加操作，当原始yolo标签文件不存在时，也可以生成新的yolo标签文件。

操作

1. COCO上的yolov4预训练模型，见56服务器上/ssd01/wanghuijiao/pose_detector02/yolov4.weights，需要自行准备/path/to/your_test_data.data文件，此文件格式为：

   classes = 80 # (yolov4.weights 模型对应检测类别)
   train   = /ssd01/wanghuijiao/dataset/crowdhuman/crowdhuman_person_head/train.txt # 训练集路径（存放图片绝对路径，且图片和txt标签文件在同一文件夹）
   valid   = /ssd01/wanghuijiao/dataset/crowdhuman/crowdhuman_person_head/val.txt # 测试集路径（存放图片绝对路径，且图片和txt标签文件在同一文件夹）
   names   = /ssd01/wanghuijiao/dataset/Obj365/coco_classes.txt # yolov4.weights 模型对应检测类别名称，COCO就直接用这个路径吧
   backup  = yolov4_human_head_detector_cfg/backup/ # 训练时权重文件输出路径，测试时这个不用管

   用以下命令对测试集生成json结果，结果默认为/ssd01/wanghuijiao/pose_detector02/results/coco_results.json:

   cd /ssd01/wanghuijiao/pose_detector02
   ./darknet detector valid \
       /path/to/your_test_data.data \
       /ssd01/wanghuijiao/pose_detector02/cfg_init/yolov4.cfg \
       /ssd01/wanghuijiao/pose_detector02/yolov4.weights \

2. 将上述测试集结果/ssd01/wanghuijiao/pose_detector02/results/coco_results.json转换并与原本的标签文件合并。/ssd01/wanghuijiao/dataset/Summary/yolor_json2yolo.py脚本内容请根据需求自行更改。

   import json
   import os
   import cv2
   from cv2 import data
   import ast
   import tqdm
   import argparse
   def convert(size, box):
       # box: x, y, w, h 绝对坐标
       dw = 1. / (size[0])
       dh = 1. / (size[1])
       x_c = box[0] + box[2] / 2.0
       y_c = box[1] + box[3] / 2.0
       w = box[2]
       h = box[3]
       # 输出相对坐标 x_center, y_center, w, h
       x_c = x_c * dw
       w = w * dw
       y_c = y_c * dh
       h = h * dh
       return (x_c, y_c, w, h)
   parser = argparse.ArgumentParser()
   parser.add_argument('--json_path',type=str ,default='./v1.0/coco_head_results.json', help="which save the output file and classes")
   parser.add_argument('--label_path_dir',type=str ,default='', help="which save the output file and classes")
   arg = parser.parse_args()
   json_path = arg.json_path
   threshold = 0.5
   label_path_src = '/ssd01/wanghuijiao/dataset/human_head_vehicle_RGB/Open_imagesV6/v2.0/labels' # 原始标签路径，请根据需求自行更改
   label_path_dir = arg.label_path_dir # '/ssd01/wanghuijiao/dataset/human_head_vehicle_RGB/Open_imagesV6/v2.0/labels' # 生成的新标签文件放置的路径，请根据需求自行更改
   img_path = '/ssd01/wanghuijiao/dataset/Open_imagesV6/person_car/images'
   # crowdhuman head detector
   # category_ids = {'0': 'head'} # '1': human
   # COCO 80 yolov4-official
   category_ids ={1: "person", 3: "car", 6: "bus", 7: "train", 8: "truck"} # 这个可以筛选json的目标标签，比如json有80类结果，但只需要1，3，6，7，8这些类。
   # 0:person, 1:car, 2:head
   with open(json_path, 'r') as f:
       for k, line in tqdm(enumerate(f.readlines())):
           if k == 0 or k == len(f.readlines()):
               continue
           line = line.strip()[:-1]
           # print(line)
           data = ast.literal_eval(line)
           if data['score'] > threshold:
               print(data['bbox'])
               image_id = data['image_id']
               category_id = data['category_id']
               if category_id not in category_ids.keys():
                   continue
               # print('category_id: ', category_id)
               bbox = data['bbox']
               bbox[0] = float(bbox[0])
               bbox[1] = float(bbox[1])
               bbox[2] = float(bbox[2])
               bbox[3] = float(bbox[3])
               img_file = os.path.join(img_path, f'{image_id}.jpg')
               img = cv2.imread(img_file)
               label_file_src = os.path.join(label_path_src, f'{image_id}.txt')
               img_height = img.shape[0]
               img_width = img.shape[1]        
               label_file_dir = os.path.join(label_path_dir, f'{image_id}.txt')
               with open(label_file_dir, 'a+') as f:
                   head_box = convert([img_width, img_height], bbox)
                   content = str('1') + ' ' + str(head_box[0])+ ' ' + str(head_box[1]) + ' ' + \
                                       str(head_box[2]) + ' ' + str(head_box[3]) + '\n' # 这里是生成的目标类别的新标签，请根据需求自行更改
                   # print(content)
                   f.write(content)