ai-labelImg-double-checker

Unless you are perfect - having some way to double check your classifications from LabelImg is a must. I didn't find anything so I made one - it is super simple. It extracts all images with the file name .check.1.MyClass1Name at the end. Then you can look at them all side by side and see how you did with a simple Explorer or Finder window filtering each class one at a time. I am sadly not perfect I found out.

Here is what it does a) for every .png/.jpg with a .txt that isn't empty it creates a sub-image clip just like you marked in LabelImg (or the like) in YOLO format (other format to come) b) It tallies all the samples for each class c) it makes a full-path list of images that were actually processed (leaving the empty .txt file items OUT of the final list) - this is used later on in the YOLO setup anyway - and I am NOT going to type them all out by hand :-) d) If something goes wrong it tells you along the way. If all goes well there is a % progress is all. e) It prints a summary at the bottom - of how many images per class - where the output directory is in case you forget or are lazy like me

Right now (cause I really got to crack on!) it only does YOLOv2/3 format - I am sure I will modify it (or someone else can) for Pascal VOC format. With YOLO, for heaven's sake remember that the x,y is the center of the square.

install

Well it is one file so just ensure you have opencv (a reminder that you cannot just pip install cv2 - you have to type pip install opencv-python and then import cv2

instructions

So if we make an image extractor - we can quickly check that our LabelImg efforts worked.

Set the sourcedir variable to the proper directory. change diffCheckDirExt if you want photos saved in another peer directory to your images directory.

It looks for jpg and pngs and if there is a corresponding .txt file it iterates through every line and makes a file called .check.X.<yourclassname>.jpg (or .png). X could be the class number starting from 0 in a file called classes.txt that needs to be along side your images - its the same class file you would use for YOLO or the like.

You have to have a classes.txt file that would have come from your LabelImg efforts anyway. I assume it is in the same directory as the images - which might not be the case for training - but I wanted to keep this simple.

code

import cv2
import numpy as np 
import pandas as pd 



sourcedir = "D:\\img"
diffCheckDirExt = ".check"
classDef = {}
classDefCount = {}
totalExtractions = 0
totalValidExtractions = 0 

def loadClasses():
    cf =  os.path.join(sourcedir,"classes.txt")
    #print(cf)
    if not os.path.exists(cf):  
        print("classes.txt file missing. Make one or add one" + cf)
        return False
    #print(cf)
    classes = pd.read_csv(cf, names = ['class'])
    print (classes)
    if not classes.count:
        return False
    
    i = 0

    for index, row in classes.iterrows(): 
        classDef[index] = row['class']
        classDefCount[index] = 0

    classes['countItems'] = 0
    #print (classDef)
    

    return True

def makeCheckDir():
    if not os.path.exists(sourcedir + diffCheckDirExt):
        os.mkdir(sourcedir + diffCheckDirExt)
    return

def preDeleteFilesOfType(intype):
    test = os.listdir(sourcedir + diffCheckDirExt)

    for item in test:
        if (item.endswith(".png") or item.endswith(".jpg")) and item.find(".check.")>=0:
            os.remove(os.path.join(sourcedir + diffCheckDirExt, item))
        if item == "file_list.txt":
            os.remove(os.path.join(sourcedir + diffCheckDirExt, item))

def handlefile(file):
    global totalExtractions
    global totalValidExtractions
    #check for .txt file with the same name
    filename = os.path.join(sourcedir,file)
    pre, ext = extension = os.path.splitext(file)
    txtFile = os.path.join(sourcedir,pre + ".txt")
    if not os.path.exists(txtFile):
        print ("Image.txt file does not exist"+txtFile)
        return
    
    tf = pd.read_csv(txtFile,delim_whitespace=True, names = ['class','xmid','ymid','w','h'])

    im = cv2.imread(filename)
    if im is None:
        print("Can't load image: " + filename)
        return
    
    for index, row in tf.iterrows():
        if index == 0:
            file_object = open(os.path.join(sourcedir+diffCheckDirExt,'file_list.txt'), 'a')
            halfpath = filename
            file_object.write(halfpath + "\n")
            file_object.close()

        totalExtractions = totalExtractions  + 1
        if exportFile(index, row, pre,ext,im):
            classDefCount[int(row['class'])] = classDefCount[int(row['class'])] + 1 #the index is numeric version of the class
            totalValidExtractions = totalValidExtractions + 1


    
            

def exportFile(index, row , pre, ext, im):

    imh = im.shape[0]-1
    imw = im.shape[1]-1
    
    #print('Image',imw, imh, str(round(row['class'])))
    x1 = round(((row['xmid'] - row['w']/2))*imw)
    x2 = round(((row['xmid'] + row['w']/2))*imw)
    y1 = round(((row['ymid'] - row['h']/2))*imh)
    y2 = round(((row['ymid'] + row['h']/2))*imh)
    #print('Image',imw, imh, x1,y1,x2,y2)
    im2 = im[y1:y2,x1:x2]
    #if im2.empty():
    #    print("Image is empty")
    #    return 

    outfn = classDef[round(row['class'])]
    outFile = os.path.join(sourcedir + diffCheckDirExt,pre + ".check."+ str(index) + "." + outfn+ext)
    #print("Writing:"+outFile)

    sh2 = im2.shape
    if sh2[0] == 0 or sh2[1] == 0:
        print ("Can't write (img empty):" +outFile)
        return False

    if not cv2.imwrite(outFile, im2):
        print ("Can't write: " + outFile)
        return False
    return True

def mainRoutine():
    filename = "testing.jpg" # was used for testing

    if not loadClasses():
        sys.exit()
    makeCheckDir()
    preDeleteFilesOfType(['.png','.jpg'])

    #handlefile(filename)
    filess = os.listdir(sourcedir)
    filessCount = len(filess)
    print (str(filessCount) + " images")
    i=0.0001
    for file in filess:
        i = i + 1
        if file.endswith(".jpg") or file.endswith(".png"):
            percentComplete = i / filessCount * 100.0
            #print(os.path.join(sourcedir, file))
            print(str(round(percentComplete)) + '% ' + os.path.join(sourcedir, file), end='\r')
            handlefile(file)
    print ("100% Complete")
    print("")
    printSummary()

def printSummary():
    print("Counts")
    print("---------------------------------------")
    for i in range(len(classDef)):
        print ("Class " + str(i) + ": " + classDef[i] + ":" + str(classDefCount[i]))
    print("")
    print ("Total Successful/Valid Extractions: " + str( round(totalValidExtractions/totalExtractions*100) )+ "%")
    print ("A reminder that your check folder is: " + sourcedir + diffCheckDirExt)
    print ("There is also a full-path files_list.txt file there ")
    
mainRoutine()