Question about elastic transform

[kronee0516]

Hi,

I am doing some training on yolov5
and I want to use albumentations to generate more data.
but i notice that the elastic transform may duplicate my origin image
and the single processed box included the all duplicated image, which is not the effect i want

may i have a way to avoid this happen?

[ternaus]

Could you please provide a code snippet that will allow tp reproduce the issue, with an example image?

[kronee0516]

the elastic jpg is the image that may produce with my code provided
actually i just want to make the license plate like viewing in different angle
elastic transform seems too slow. is affine the better option for my need?

ALABAMA_train_0.txt

import cv2
import albumentations as A
import random


def visualizeBbox(img,bbox):
    dh, dw, _ = img.shape
    #fl = open( 'asian-baby-in-living-room-picture-id878706102_jpg.rf.7000326f938e1bf2de0828b639c0ce4e.txt', 'r')
    #data = fl.readlines()
    #fl.close()


    # Split string to float
    for box in bbox:
        x, y, w, h = box

        # Taken from https://github.com/pjreddie/darknet/blob/810d7f797bdb2f021dbe65d2524c2ff6b8ab5c8b/src/image.c#L283-L291
        # via https://stackoverflow.com/questions/44544471/how-to-get-the-coordinates-of-the-bounding-box-in-yolo-object-detection#comment102178409_44592380
        l = int((x - w / 2) * dw)
        r = int((x + w / 2) * dw)
        t = int((y - h / 2) * dh)
        b = int((y + h / 2) * dh)

        if l < 0:
            l = 0
        if r > dw - 1:
            r = dw - 1
        if t < 0:
            t = 0
        if b > dh - 1:
            b = dh - 1

        cv2.rectangle(img, (l, t), (r, b), (0, 0, 255), 1)

    cv2.imshow("",img)
    cv2.waitKey()

def read_box(path):
    ll=[]
    fl = open(path, 'r')
    data = fl.readlines()
    fl.close()
    for datas in data:
        _, x, y, w, h = map(float, datas.split(' '))
        ll.append([x,y,w,h])
    return ll

def save_box(filename,boxs,cl):
    tempstr=''
    with open(filename,'w') as f:
        for box in boxs:
            [x,y,w,h]=box
            tempstr = tempstr+  str(cl)+' '+str(x)+' '+str(y)+' '+str(w)+' '+str(h)
            if boxs[-1]!=box:
                tempstr=tempstr+'\n'
        f.write(tempstr)
        f.close()

def buildTF(resize_only):

    imp_list=[]

    randsize = random.randint(100, 224)
    print(f'randsize : {randsize}')
    imp_list.append(A.LongestMaxSize(max_size=randsize, interpolation=1))
    imp_list.append(A.PadIfNeeded(position='random',min_height=640, min_width=640, border_mode=0, value=(0, 0, 0)))
    imp_list.append(A.ElasticTransform(alpha=1,sigma=50,alpha_affine=90,p=1))
    #,same_dxdy=True

    transforml = A.Compose(
            imp_list
        ,bbox_params=A.BboxParams(format="yolo",label_fields="")
    )
    return transforml


img = "ALABAMA_train_0.jpg"
label = "ALABAMA_train_0.txt"
imgtemp = cv2.imread(img)
tfx = buildTF(False)#rebuild auglist for cutout random
augmentations = tfx(image=imgtemp,bboxes=read_box(label))
augimg = augmentations["image"]
visualizeBbox(augimg,augmentations['bboxes'])

[kronee0516]

another question
is there any function that like padifneed but pad with other image instead of image itself or plain color edges?