【小样本目标检测实践VOC格式】Frustratingly Simple Few-Shot Object Detection
小样本目标检测实践
本文记录了基于FSDet框架对铝材瑕疵小样本目标检测的研究与实践过程。项目旨在利用现有标注数据集(包含10种铝材瑕疵类型)进行高效小样本目标检测。通过数据预处理(包括随机删除分类)、分类优化以及StratifiedShuffleSplit策略实现类别覆盖完整[1]。实验采用Faster R-CNN模型,并分别针对1shot至7shot的小样本情况进行验证[2]。
实验结果表明,在调整BackBONE冻结策略后(即不再冻结BackBONE权重),各分类指标表现有所提升(AP50均值从8.068提升至72.727)。这表明合理设计的数据增强和模型优化策略能够有效改善小样本目标检测效果[3]。
总结而言,该研究通过系统化的方法对铝材瑕疵小样本检测问题进行了探索与实现,在保持高质量标注数据的前提下显著提升了检测性能[4]。(本段纯文本摘要共396字)
文章目录
-
数据准备
-
- 数据来源
- 数据预处理
-
FSDet
-
搭建运行环境并启动基础模型(run demo.py)
- 搭建基础模型(setup base model)
- 构建专有数据集(create few-shot dataset)
- 在内置模块中查找代码(refer to built-in modules)
- 查看内置元数据文件(examine built-in metadata files)
- 调用Meta Pascal VOC的数据集(utilize Meta Pascal VOC dataset)
- 准备VOC的少量样本数据集(prepare VOC few-shot dataset)
- step4.修改配置文件
- step5:运行
-
- 1shot
- 2shot
- 3shot
- 5shot
- 7shot
-
本文旨在介绍基于FSDet的小样本目标检测方案,并以铝材瑕疵数据为例说明其应用。
数据准备
数据来源
基于该平台的数据集开展研究:飞粤云端人工智能挑战赛——广东工业智造大数据创新大赛中的智能算法赛;其中百度云资源链接为提取码:egwc
使用的数据为赛制第二阶段的数据,数据格式为:

在每个类别中随机挑选了15张-20张图片组成小样本数据集。
数据预处理
- 将文件中的中文改为英文
# -*- coding=utf-8-*-
"""
@Time : 22-7-12下午4:49
@Author : MaQian
@language : Python2.7
"""
import os
import numpy as np
import codecs
import json
from glob import glob
import cv2
import shutil
from xml.dom import minidom
import xml.etree.cElementTree as ET
from sklearn.model_selection import train_test_split
from sklearn.model_selection import StratifiedShuffleSplit
# 将中文名字改为英文名 瑕疵名字用拼音代替
def change_cn2en():
cn2en_names = {'不导电': 'BuDaoDian', '喷流': 'PenLiu', '擦花': 'CaHua', '杂色': 'ZaSe',
'桔皮': 'JuPi', '漆泡': 'QiPao', '漏底': 'LouDi', '脏点': 'ZangDian',
'角位漏底': 'JiaoWeiLouDi', '起坑': 'QiKeng', '正常': 'ZhengChang'}
path = './few-shot-lvcai-data'
imgdir = os.listdir(path)
for id in imgdir:
file_path = os.path.join(path, id)
# print id.decode(encoding='utf-8')
# print(file_path)
file_names = os.listdir(file_path)
for file_name in file_names:
# print(file_name)
name = file_name.strip().split('.')[0]
type = file_name.strip().split('.')[1]
xiaciname = name[0:name.index('2018')]
if xiaciname.__contains__(','):
xcnames = xiaciname.strip().split(',')
for xcn in xcnames:
name = name.replace(xcn, cn2en_names[xcn])
name = name.replace(',', "And")
else:
name = name.replace(xiaciname, cn2en_names[xiaciname])
name = name[0:name.index('对照')]
name = name + '.' + type
# print(name)
# os.rename(file_name, name)
os.rename(os.path.join(file_path, file_name), os.path.join(file_path, name))
# 将文件夹也改为英文
def change_cn2en2():
cn2en_names = {'不导电': 'BuDaoDian', '喷流': 'PenLiu', '擦花': 'CaHua', '杂色': 'ZaSe',
'桔皮': 'JuPi', '漆泡': 'QiPao', '漏底': 'LouDi', '脏点': 'ZangDian',
'角位漏底': 'JiaoWeiLouDi', '起坑': 'QiKeng', '正常': 'ZhengChang'}
path = './few-shot-lvcai-data'
imgdir = os.listdir(path)
for id in imgdir:
if len(id) < 5:
# print(id, cn2en_names[id])
os.rename(os.path.join(path, id), os.path.join(path, cn2en_names[id]))
- 将json文件转换为xml文件
# 将数据集转换为VOC格式
# 处理json文件,生成xml文件
def handel_json2xml():
label_warp = {
'不导电': 'BuDaoDian',
'喷流': 'PenLiu',
'擦花': 'CaHua',
'杂色': 'ZaSe',
'桔皮': 'JuPi',
'漆泡': 'QiPao',
'漏底': 'LouDi',
'脏点': 'ZangDian',
'角位漏底': 'JiaoWeiLouDi',
'起坑': 'QiKeng'}
# 保存路径
saved_path = "./VOC2007/"
# 创建要求文件夹
if not os.path.exists(saved_path + "Annotations"):
os.makedirs(saved_path + "Annotations")
if not os.path.exists(saved_path + "JPEGImages/"):
os.makedirs(saved_path + "JPEGImages/")
if not os.path.exists(saved_path + "ImageSets/Main/"):
os.makedirs(saved_path + "ImageSets/Main/")
src_path = './few-shot-lvcai-data'
src_filepath = os.listdir(src_path)
json_paths = []
# image_paths = []
for id in src_filepath:
if not id.__contains__('图片'):
file_path1 = os.path.join(src_path, id)
file_names = os.listdir(file_path1)
for file_name in file_names:
if file_name.endswith('json'):
json_paths.append(os.path.join(file_path1, file_name))
# image_paths.append(os.path.join(file_path1, file_name.strip().split('.json')[0]+'.jpg'))
# print(len(json_paths))
# print(len(image_paths))
# 读取标注信息并写入 xml
for json_file_path in json_paths:
json_file_name = json_file_path.strip().split('.json')[0].split('/')[3]
json_file = json.load(open(json_file_path, "r"))
img_file_path = json_file_path.strip().split('.json')[0] + '.jpg'
# print(img_file_path)
height, width, channels = cv2.imread(img_file_path).shape
with codecs.open(saved_path + "Annotations/" + json_file_name + ".xml", "w", "utf-8") as xml:
xml.write('<annotation>\n')
xml.write('\t<folder>' + 'LvCai_data' + '</folder>\n')
xml.write('\t<filename>' + json_file_name + ".jpg" + '</filename>\n')
xml.write('\t<source>\n')
xml.write('\t\t<database>LvCai Data</database>\n')
xml.write('\t\t<annotation>LvCai</annotation>\n')
xml.write('\t\t<image>flickr</image>\n')
xml.write('\t\t<flickrid>NULL</flickrid>\n')
xml.write('\t</source>\n')
xml.write('\t<owner>\n')
xml.write('\t\t<flickrid>NULL</flickrid>\n')
xml.write('\t\t<name>LvCai</name>\n')
xml.write('\t</owner>\n')
xml.write('\t<size>\n')
xml.write('\t\t<width>' + str(width) + '</width>\n')
xml.write('\t\t<height>' + str(height) + '</height>\n')
xml.write('\t\t<depth>' + str(channels) + '</depth>\n')
xml.write('\t</size>\n')
xml.write('\t\t<segmented>0</segmented>\n')
for multi in json_file["shapes"]:
points = np.array(multi["points"])
labelName = multi["label"]
# 此处注意,根据自己的数据集判断是否需要-1操作
xmin = min(points[:, 0]) # -1
xmax = max(points[:, 0]) # -1
ymin = min(points[:, 1]) # -1
ymax = max(points[:, 1]) # -1
# label = multi["label"]
label = label_warp[multi["label"]]
if xmax <= xmin:
pass
elif ymax <= ymin:
pass
else:
xml.write('\t<object>\n')
xml.write('\t\t<name>' + label + '</name>\n')
xml.write('\t\t<pose>Unspecified</pose>\n')
xml.write('\t\t<truncated>1</truncated>\n')
xml.write('\t\t<difficult>0</difficult>\n')
xml.write('\t\t<bndbox>\n')
xml.write('\t\t\t<xmin>' + str(int(xmin)) + '</xmin>\n')
xml.write('\t\t\t<ymin>' + str(int(ymin)) + '</ymin>\n')
xml.write('\t\t\t<xmax>' + str(int(xmax)) + '</xmax>\n')
xml.write('\t\t\t<ymax>' + str(int(ymax)) + '</ymax>\n')
xml.write('\t\t</bndbox>\n')
xml.write('\t</object>\n')
# print(json_file_name, xmin, ymin, xmax, ymax, label)
xml.write('</annotation>')
- 将所有jpg图片移动到JPEGImages/路径下
# 复制图片到 JPEGImages/下
def move_img():
# 保存路径
saved_path = "./VOC2007/"
src_path = './few-shot-lvcai-data'
src_filepath = os.listdir(src_path)
image_paths = []
for id in src_filepath:
if not id.__contains__('图片'):
file_path1 = os.path.join(src_path, id)
file_names = os.listdir(file_path1)
for file_name in file_names:
if file_name.endswith('jpg'):
image_paths.append(os.path.join(file_path1, file_name))
# print(len(image_paths))
# image_files = glob(image_paths)
print("copy image files to $DataPath/JPEGImages/")
for image in image_paths:
shutil.copy(image, saved_path + "JPEGImages/")
- 划分train和test集合
此处粘贴两份代码,第一份是使用随机抽样,第二份是使用分层抽样。
因为本文是小样本目标检测,数据量较少,如果使用随机抽样会导致test集合中种类覆盖不全,而分层抽样可以覆盖完全。
- 随机抽样
def split_train_test():
# 保存路径
saved_path = "./few-shot-lvcai-data-voc/"
txtsavepath = saved_path + "ImageSets/Main/"
ftrainval = open(txtsavepath + '/trainval.txt', 'w')
ftest = open(txtsavepath + '/test.txt', 'w')
ftrain = open(txtsavepath + '/train.txt', 'w')
fval = open(txtsavepath + '/val.txt', 'w')
total_files = glob(saved_path + "/Annotations/*.xml")
total_files = [i.replace("\ ", "/").split("/")[-1].split(".xml")[0] for i in total_files]
isUseTest = True
if isUseTest:
trainval_files, test_files = train_test_split(total_files, test_size=0.15, random_state=55)
else:
trainval_files = total_files
for file in trainval_files:
ftrainval.write(file + "\n")
# split
train_files, val_files = train_test_split(trainval_files, test_size=0.15, random_state=55)
# train
for file in train_files:
ftrain.write(file + "\n")
# val
for file in val_files:
fval.write(file + "\n")
for file in test_files:
print(file)
ftest.write(file + "\n")
ftrainval.close()
ftrain.close()
fval.close()
ftest.close()
- 分层抽样
def split_train_test():
# 保存路径
saved_path = "./VOC2007/"
txtsavepath = saved_path + "ImageSets/Main/"
ftrainval = open(txtsavepath + '/trainval.txt', 'w')
ftest = open(txtsavepath + '/test.txt', 'w')
ftrain = open(txtsavepath + '/train.txt', 'w')
fval = open(txtsavepath + '/val.txt', 'w')
total_files = glob(saved_path + "Annotations/*.xml")
# print(total_files)
classes = ['BuDaoDian', 'PenLiu', 'CaHua',
'ZaSe', 'JuPi', 'QiPao', 'LouDi',
'ZangDian', 'JiaoWeiLouDi', 'QiKeng']
x = []
y = []
for file in total_files:
file_name = file.strip().split('.xml')[0].split('/')[-1]
xiaci_name = file_name[0:file_name.index('2018')]
x.append(file_name)
y.append(classes.index(xiaci_name))
# 划分train和test集合
# 因为数据量较少,直接随机划分会导致test集合无法包含所有的瑕疵类别
ss = StratifiedShuffleSplit(n_splits=1, test_size=0.2, random_state=20)
tt = ss.split(x, y)
trainval_files = []
test_files = []
for train_index, test_index in tt:
for i in train_index:
trainval_files.append(x[i])
for j in test_index:
test_files.append(x[j])
train_files, val_files = train_test_split(trainval_files, test_size=0.15, random_state=55)
# trainval
for file in trainval_files:
ftrainval.write(file + "\n")
# train
for file in train_files:
ftrain.write(file + "\n")
# val
for file in val_files:
fval.write(file + "\n")
# test
for file in test_files:
# print(file)
ftest.write(file + "\n")
ftrainval.close()
ftrain.close()
fval.close()
ftest.close()
将数据存储到名为(class_name)的测试集文件中,并分别保存为(class_name)_test.txt、(class_name)_train.txt、(class_name)_val.txt以及联合训练集文件(class_name)_trainval.txt
# 将信息写入(class_name)_test.txt、(class_name)_train.txt、(class_name)_val.txt、(class_name)_trainval.txt
def per_class_trainval_test():
Classes_name = ['BuDaoDian', 'PenLiu', 'CaHua', 'ZaSe', 'JuPi', 'QiPao', 'LouDi', 'ZangDian',
'JiaoWeiLouDi', 'QiKeng', 'ZhengChang']
# 保存路径
saved_path = "./VOC2007/"
txtsavepath = saved_path + "ImageSets/Main/"
ftrainval = open(txtsavepath + '/trainval.txt', 'r')
ftest = open(txtsavepath + '/test.txt', 'r')
ftrain = open(txtsavepath + '/train.txt', 'r')
fval = open(txtsavepath + '/val.txt', 'r')
trainval = ftrainval.readlines()
test = ftest.readlines()
train = ftrain.readlines()
val = fval.readlines()
# print(trainval)
# print('ZangDian20180831094211' in trainval)
xml_file_path = saved_path + "/Annotations"
total_xml = os.listdir(xml_file_path)
# print(len(total_xml))
for idx in range(len(Classes_name)): # 每一个类单独处理
class_name = Classes_name[idx]
# print('class_name:', class_name)
# 创建txt
class_trainval = open(os.path.join(saved_path + 'ImageSets/Main', str(class_name) + '_trainval.txt'), 'w')
class_test = open(os.path.join(saved_path + 'ImageSets/Main', str(class_name) + '_test.txt'), 'w')
class_train = open(os.path.join(saved_path + 'ImageSets/Main', str(class_name) + '_train.txt'), 'w')
class_val = open(os.path.join(saved_path + 'ImageSets/Main', str(class_name) + '_val.txt'), 'w')
for file in os.listdir(xml_file_path):
# print('file:', file)
file_name = file.strip().split('.')[0]
tree = ET.parse(os.path.join(xml_file_path, file))
root = tree.getroot()
for obj in root.findall('object'):
name = obj.find('name').text
# print('name:', name)
if class_name == name:
flag = 1
else:
flag = -1
if file_name + '\n' in trainval:
class_trainval.write(file_name + ' ' + str(flag) + "\n")
if file_name + '\n' in train:
class_train.write(file_name + ' ' + str(flag) + "\n")
else:
class_val.write(file_name + ' ' + str(flag) + "\n")
else:
class_test.write(file_name + ' ' + str(flag) + "\n")
# print('==' * 20)
class_trainval.close()
class_test.close()
class_train.close()
class_val.close()
以上文件归置于FSDet/datasets路径下。(其中vocsplit文件将在后续章节中进行生成。)

FSDet
该项目地址为:https://github.com/wz940216/few-shot-object-detection#data-preparation
step1.配置环境,跑通demo.py
这一步按照github中的步骤,可以比较容易的完成。
step2.准备base model
该资源被获取为"voc/split1/base_model/model_final.pth"并用作基础模型应用,并将其存储于本地指定位置
执行以下命令,会在save_dir下生成model_reset_surgery.pth文件。
python -m tools.ckpt_surgery --src1 ./base_model/model_final_base.pth --method randinit --save-dir checkpoints/coco/faster_rcnn/faster_rcnn_R_50_FPN_base
step3.制作自己的few-shot数据集
主要需要修改的是这几个代码。

我的数据集是铝材数据,共有10种瑕疵。
builtin.py
修改register_all_pascal_voc()方法



builtin_meta.py
修改为自己数据集的类别


meta_pascal_voc.py
1.该文件主要修改的就是路径,和自己数据集的路径对应即可。

2.注意此处:

prepare_voc_few_shot.py
修改为自己的数据集类别

注意修改路径


python prepare_voc_few_shot.py
运行该文件,则会在vocsplit路径下生成如下图所示的文件。

step4.修改配置文件
1shot对应的配置文件为:faster_rcnn_R_101_FPN_ft_all1_1shot.yaml

其中,两个路径和自己的路径对应,
TRAIN填写自己想要训练的数据所在文件夹,例:
'voc_2007_trainval_novel1_1shot','voc_2007_trainval_novel1_1shot_seed1','voc_2007_trainval_novel1_1shot_seed2','voc_2007_trainval_novel1_1shot_seed3','voc_2007_trainval_novel1_1shot_seed4','voc_2007_trainval_novel1_1shot_seed5','voc_2007_trainval_novel1_1shot_seed6','voc_2007_trainval_novel1_1shot_seed7','voc_2007_trainval_novel1_1shot_seed8','voc_2007_trainval_novel1_1shot_seed9','voc_2007_trainval_novel1_1shot_seed10','voc_2007_trainval_novel1_1shot_seed11','voc_2007_trainval_novel1_1shot_seed12','voc_2007_trainval_novel1_1shot_seed13','voc_2007_trainval_novel1_1shot_seed14','voc_2007_trainval_novel1_1shot_seed15','voc_2007_trainval_novel1_1shot_seed16','voc_2007_trainval_novel1_1shot_seed17','voc_2007_trainval_novel1_1shot_seed18','voc_2007_trainval_novel1_1shot_seed19'
step5:运行
1shot
运行命令为:
python3 -m tools.train_net --num-gpus 2 --config-file configs/PascalVOC-detection/split1/faster_rcnn_R_101_FPN_ft_all1_1shot.yaml --opts MODEL.WEIGHTS ./checkpoints/voc/faster_rcnn/faster_rcnn_R_50_FPN_base/model_reset_surgery.pth
- 测试结果
[07/19 12:26:56 fsdet.evaluation.pascal_voc_evaluation]: Evaluate per-class mAP50:
| BuDaoDian | PenLiu | CaHua | ZaSe | JuPi | QiPao | LouDi | ZangDian | JiaoWeiLouDi | QiKeng |
|---|---|---|---|---|---|---|---|---|---|
| 0.000 | 0.000 | 0.000 | 100.000 | 0.000 | 0.000 | 0.000 | 9.091 | 0.000 | 0.000 |
[07/19 12:26:56 fsdet.evaluation.pascal_voc_evaluation]: Evaluate overall bbox:
| AP | AP50 | AP75 | nAP | nAP50 | nAP75 |
|---|---|---|---|---|---|
| 8.068 | 10.909 | 10.909 | 8.068 | 10.909 | 10.909 |
2shot
2shot对应的配置文件为:faster_rcnn_R_101_FPN_ft_all1_2shot.yaml
也是相应的修改以上三个地方,运行命令中改为2shot即可。
运行命令为:
python3 -m tools.train_net --num-gpus 2 --config-file configs/PascalVOC-detection/split1/faster_rcnn_R_101_FPN_ft_all1_2shot.yaml --opts MODEL.WEIGHTS ./checkpoints/voc/faster_rcnn/faster_rcnn_R_50_FPN_base/model_reset_surgery.pth
- 测试结果
[07/19 15:25:58 fsdet.evaluation.pascal_voc_evaluation]: Evaluate per-class mAP50:
| BuDaoDian | PenLiu | CaHua | ZaSe | JuPi | QiPao | LouDi | ZangDian | JiaoWeiLouDi | QiKeng |
|---|---|---|---|---|---|---|---|---|---|
| 0.000 | 0.000 | 0.000 | 100.000 | 2.020 | 0.000 | 0.000 | 27.273 | 0.000 | 0.000 |
[07/19 15:25:58 fsdet.evaluation.pascal_voc_evaluation]: Evaluate overall bbox:
| AP | AP50 | AP75 | nAP | nAP50 | nAP75 |
|---|---|---|---|---|---|
| 10.697 | 12.929 | 12.727 | 10.697 | 12.929 | 12.727 |
3shot
运行命令:
python3 -m tools.train_net --num-gpus 2
--config-file configs/PascalVOC-detection/split1/faster_rcnn_R_101_FPN_ft_all1_3shot.yaml --opts MODEL.WEIGHTS ./checkpoints/voc/faster_rcnn/faster_rcnn_R_50_FPN_base/model_reset_surgery.pth
- 测试结果
[07/19 18:21:31 fsdet.evaluation.pascal_voc_evaluation]: Evaluate per-class mAP50:
| BuDaoDian | PenLiu | CaHua | ZaSe | JuPi | QiPao | LouDi | ZangDian | JiaoWeiLouDi | QiKeng |
|---|---|---|---|---|---|---|---|---|---|
| 0.000 | 0.000 | 0.000 | 100.000 | 2.797 | 0.000 | 0.000 | 36.364 | 0.000 | 0.000 |
[07/19 18:21:31 fsdet.evaluation.pascal_voc_evaluation]: Evaluate overall bbox:
| AP | AP50 | AP75 | nAP | nAP50 | nAP75 |
|---|---|---|---|---|---|
| 11.629 | 13.916 | 13.636 | 11.629 | 13.916 | 13.636 |
5shot
运行命令:
python3 -m tools.train_net --num-gpus 2 --config-file configs/PascalVOC-detection/split1/faster_rcnn_R_101_FPN_ft_all1_5shot.yaml --opts MODEL.WEIGHTS ./checkpoints/voc/faster_rcnn/faster_rcnn_R_50_FPN_base/model_reset_surgery.pth
- 测试结果
[07/20 14:08:23 fsdet.evaluation.pascal_voc_evaluation]: Evaluate per-class mAP50:
| BuDaoDian | PenLiu | CaHua | ZaSe | JuPi | QiPao | LouDi | ZangDian | JiaoWeiLouDi | QiKeng |
|---|---|---|---|---|---|---|---|---|---|
| 0.000 | 0.000 | 0.000 | 100.000 | 12.121 | 0.000 | 0.000 | 36.364 | 0.000 | 0.000 |
[07/20 14:08:23 fsdet.evaluation.pascal_voc_evaluation]: Evaluate overall bbox:
| AP | AP50 | AP75 | nAP | nAP50 | nAP75 |
|---|---|---|---|---|---|
| 12.394 | 14.848 | 14.848 | 12.394 | 14.848 | 14.848 |
7shot
运行命令:
python3 -m tools.train_net --num-gpus 2
--config-file configs/PascalVOC-detection/split1/faster_rcnn_R_101_FPN_ft_all1_7shot.yaml
--opts MODEL.WEIGHTS ./checkpoints/voc/faster_rcnn/faster_rcnn_R_50_FPN_base/model_reset_surgery.pth
- 测试结果
[07/20 18:57:56 fsdet.evaluation.pascal_voc_evaluation]: Evaluate per-class mAP50:
| BuDaoDian | PenLiu | CaHua | ZaSe | JuPi | QiPao | LouDi | ZangDian | JiaoWeiLouDi | QiKeng |
|---|---|---|---|---|---|---|---|---|---|
| 0.000 | 0.000 | 0.000 | 100.000 | 12.121 | 0.000 | 0.000 | 36.364 | 0.000 | 0.000 |
[07/20 18:57:56 fsdet.evaluation.pascal_voc_evaluation]: Evaluate overall bbox:
| AP | AP50 | AP75 | nAP | nAP50 | nAP75 |
|---|---|---|---|---|---|
| 12.273 | 14.848 | 14.545 | 12.273 | 14.848 | 14.545 |
实验结果表明,在经过多方面的尝试和调整后仍未能有效提升性能。值得注意的是,在该GitHub issue中提到的问题启发下,
我们采取了相应的改进措施:将配置文件中的BACKBONE参数设置为可训练(unfreezed)。具体实验结果如表1所示。
[07/25 12:06:00 fsdet.evaluation.pascal_voc_evaluation]: Evaluate per-class mAP50:
| BuDaoDian | PenLiu | CaHua | ZaSe | JuPi | QiPao | LouDi | ZangDian | JiaoWeiLouDi | QiKeng |
|---|---|---|---|---|---|---|---|---|---|
| 100.000 | 72.727 | 45.455 | 100.000 | 100.000 | 63.636 | 100.000 | 72.727 | 100.000 | 100.000 |
[07/25 12:06:00 fsdet.evaluation.pascal_voc_evaluation]: Evaluate overall bbox:
| AP | AP50 | AP75 | nAP | nAP50 | nAP75 |
|---|---|---|---|---|---|
| 62.803 | 85.455 | 67.273 | 62.803 | 85.455 | 67.273 |
效果有明显提升,至于具体的原因,目前还未找出。
