My implementation of AutoAugmentation, RandomAugmentation and AutoAugmentation for detection with cv2.
You can download this repo to your project:
cd /your/project
git clone https://github.com/CoinCheung/AutoAugment_opencv.git
and import it directly.
If your task is classification, you can do this:
from AutoAugment_opencv.AA_classification import AutoAugment, RandomAugment
import cv2
img = cv2.imread('./AutoAugment_opencv/pic.jpg')
AA = AutoAugment()
aa_auged = AA(img)
RA = RandomAugment(N=2, M=9)
ra_auged = RA(img)If your task is object detection, you can use autoaugment for detection like this:
from AutoAugment_opencv.AA_detection import PolicyV0, PolicyV1, PolicyV2, PolicyV3
import xml.etree.ElementTree as ET
import cv2
import numpy as np
xmlpth = './pascal_format.xml'
jpgpth = './example_picture.jpg'
img = cv2.imread(jpgpth)
tree = ET.parse(xmlpth)
root = tree.getroot()
bboxes = []
for obj in root.findall('object'):
bnd_box = obj.find('bndbox')
bbox = [
int(bnd_box.find('xmin').text),
int(bnd_box.find('ymin').text),
int(bnd_box.find('xmax').text),
int(bnd_box.find('ymax').text)
]
bboxes.append(bbox)
AA_det = PolicyV0() # PolicyV1, PolicyV2, PolicyV3 are results also worth trying in the paper
aug_img, aug_bboxes = AA_det(img, bboxes)This implementation only matches PIL when the cv2 is 3.4.2, if it is 4.1.2, there will be slight difference.
If you see any errors in this codebase, please be my guest to open an issue to correct me. Thanks !!