Target large, medium and small numbers in the statistical data set
I recently saw a paper in which the information of large, medium and small targets was counted in the introduction part of the data set. Therefore, in order to obtain the statistical information of the data set, I refer to the code written by the author to count the number of large, medium and small targets in the coco data set based on tensorpack
The simplified version of the code is as follows (there are multiple parameters that need to be modified, please check the comments carefully)
# 1、统计数据集中小、中、大 GT的个数
# 2、统计某个类别小、中、大 GT的个数
# 3、统计数据集中ss、sm、sl GT的个数
import os
from pathlib import Path
import matplotlib.pyplot as plt
# 设置中文字体为微软雅黑
plt.rcParams['font.sans-serif'] = 'SimHei'
def getGtAreaAndRatio(label_dir):
"""
得到不同尺度的gt框个数
:params label_dir: label文件地址
:return data_dict: {dict: 3} 3 x {'类别':{’area':[...]}, {'ratio':[...]}}
"""
data_dict = {
}
assert Path(label_dir).is_dir(), "label_dir is not exist"
txts = os.listdir(label_dir) # 得到label_dir目录下的所有txt GT文件
for txt in txts: # 遍历每一个txt文件
with open(os.path.join(label_dir, txt), 'r') as f: # 打开当前txt文件 并读取所有行的数据
lines = f.readlines()
for line in lines: # 遍历当前txt文件中每一行的数据
temp = line.split() # str to list{5}
coor_list = list(map(lambda x: x, temp[1:])) # [x, y, w, h]
area = float(coor_list[2]) * float(coor_list[3]) # 计算出当前txt文件中每一个gt的面积
# center = (int(coor_list[0] + 0.5*coor_list[2]),
# int(coor_list[1] + 0.5*coor_list[3]))
ratio = round(float(coor_list[2]) / float(coor_list[3]), 2) # 计算出当前txt文件中每一个gt的 w/h
if temp[0] not in data_dict:
data_dict[temp[0]] = {
}
data_dict[temp[0]]['area'] = []
data_dict[temp[0]]['ratio'] = []
data_dict[temp[0]]['area'].append(area)
data_dict[temp[0]]['ratio'].append(ratio)
return data_dict
def getSMLGtNumByClass(data_dict, class_num):
"""
计算某个类别的小物体、中物体、大物体的个数
params data_dict: {dict: 3} 3 x {'类别':{’area':[...]}, {'ratio':[...]}}
params class_num: 类别 0, 1, 2
return s: 该类别小物体的个数 0 < area <= 0.5%
m: 该类别中物体的个数 0.5% < area <= 1%
l: 该类别大物体的个数 area > 1%
"""
s, m, l = 0, 0, 0
# 图片的尺寸大小 注意修改!!!
h = 960
w = 540
for item in data_dict['{}'.format(class_num)]['area']:
if item * h * w <= h * w * 0.005:
s += 1
elif item * h * w <= h * w * 0.010:
m += 1
else:
l += 1
return s, m, l
def getAllSMLGtNum(data_dict, isEachClass=False):
"""
数据集所有类别小、中、大GT分布情况
isEachClass 控制是否按每个类别输出结构
"""
S, M, L = 0, 0, 0
# 需要手动初始化下，有多少个类别就需要写多个
classDict = {
'0': {
'S': 0, 'M': 0, 'L': 0}, '1': {
'S': 0, 'M': 0, 'L': 0}, '2': {
'S': 0, 'M': 0, 'L': 0},
'3': {
'S': 0, 'M': 0, 'L': 0}}
print(classDict['0']['S'])
# range(class_num)类别数 注意修改!!!
if isEachClass == False:
for i in range(4):
s, m, l = getSMLGtNumByClass(data_dict, i)
S += s
M += m
L += l
return [S, M, L]
else:
for i in range(4):
S = 0
M = 0
L = 0
s, m, l = getSMLGtNumByClass(data_dict, i)
S += s
M += m
L += l
classDict[str(i)]['S'] = S
classDict[str(i)]['M'] = M
classDict[str(i)]['L'] = L
return classDict
# 画图函数
def plotAllSML(SML):
x = ['S:[0, 32x32]', 'M:[32x32, 96x96]', 'L:[96*96, 640x640]']
fig = plt.figure(figsize=(10, 8)) # 画布大小和像素密度
plt.bar(x, SML, width=0.5, align="center", color=['skyblue', 'orange', 'green'])
for a, b, i in zip(x, SML, range(len(x))): # zip 函数
plt.text(a, b + 0.01, "%d" % int(SML[i]), ha='center', fontsize=15, color="r") # plt.text 函数
plt.xticks(fontsize=15)
plt.yticks(fontsize=15)
plt.xlabel('gt大小', fontsize=16)
plt.ylabel('数量', fontsize=16)
plt.title('广佛手病虫害训练集小、中、大GT分布情况(640x640)', fontsize=16)
plt.show()
# 保存到本地
# plt.savefig("")
if __name__ == '__main__':
labeldir = r'E:\project\py-project\network\dataset\UA-DETRAC-G2\labels\val'
data_dict = getGtAreaAndRatio(labeldir)
# 1、数据集所有类别小、中、大GT分布情况
# 控制是否按每个类别输出结构
isEachClass = False
SML = getAllSMLGtNum(data_dict, isEachClass)
print(SML)
if not isEachClass:
plotAllSML(SML)