Albumentations：物体検出（枠）にも対応したデータ拡張ライブラリを解説

!pip install -U albumentations

pip install -U albumentations

import albumentations as A
import cv2

transform = A.Compose([
    A.RandomCrop(width=256, height=256),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
])

img = cv2.imread('/path/to/image.jpg')
img = cv2.cv2Color(img, cv2.COLOR_BGR2RGB)
transformed_img = transform(image = img)

transform = A.Compose([
    A.RandomCrop(width=256, height=256),
    A.OneOf([
      A.HorizontalFlip(),
      A.RandomBrightnessContrast(),
    ], p = 0.5)
])

result = A.ChannelShuffle(p=1)(image=img)

result = A.CLAHE(p=1)(image=img)

result = A.ColorJitter(brightness=1.0, hue=(-0.5,0.5), p=1)(image=img)

result = A.Downscale(scale_min=0.1, scale_max=0.25, interpolation=cv2.INTER_LINEAR, p=1)(image=img)

result = A.GaussNoise(var_limit=(100, 150), p=1)(image=img)

result = A.ImageCompression(quality_lower=1, quality_upper=10, p=1)(image=img)

result = A.InvertImg(p=1)(image=img)

result = A.Sharpen(alpha=(0, 1.0), p=1)(image=img)

result = A.Normalize(p=1)(image=img)

import albumentations.pytorch

result = A.pytorch.ToTensorV2()(image=img)
result['image'].shape

result = A.AdvancedBlur(
    blur_limit=(3, 15), 
    sigmaX_limit=(0.5, 3), 
    sigmaY_limit=(0.5, 3),  p=1)(image=img)

result = A.Blur()(image=img)

bboxes = [[0.01, 0.55 , 1.0,0.94]]

def clamp(n, smallest, largest):
    return max(smallest, min(n, largest))

result = A.BBoxSafeRandomCrop(erosion_rate=0.3)(image=img, bboxes=bboxes)
img2 = result['image'].copy()
for bbox in result['bboxes']:
  x1, y1, x2, y2 = bbox
  h, w, _ = img2.shape
  x1 = clamp(x1, 0, 1)
  y1 = clamp(y1, 0, 1)
  x2 = clamp(x2, 0, 1)
  y2 = clamp(y2, 0, 1)
  cv2.rectangle(img2, (int(x1*w), int(y1*h)), (int(x2*w), int(y2*h)), (0,0,255), 4)

cv2_imshow(img2)

result = A.CenterCrop(width=200, height=200)(image=img)

result = A.Crop(x_min=0, y_min=0, x_max=256, y_max=100)(image=img)

result = A.RandomCrop(width=200, height=200)(image=img)

result = A.CoarseDropout(
    max_holes= 100,
    max_height=16, 
    max_width =16, p=1)(image=img)

result = A.Cutout (
    num_holes=100, 
    max_h_size=16, max_w_size=16, 
    fill_value=(255,0,0), p=1)(image=img)

result = A.LongestMaxSize(max_size=256)(image=img)

result = A.RandomScale(scale_limit=(-1, 1), p=1)(image=img)

result = A.Resize(width=100, height=100, p=1)(image=img)

result = A.SmallestMaxSize(max_size=256)(image=img)

ref = cv2.imread("dark.png")
result = A.HistogramMatching([ref], read_fn=lambda x: x, p=1)(image=img)

result = A.RandomFog(fog_coef_lower=0.3, fog_coef_upper=1, alpha_coef=0.08, p=1)(image=img)

result = A.RandomRain(p=1)(image=img)

transform = A.Compose([
    A.RandomCrop(width=256, height=256, p=1),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
], bbox_params=A.BboxParams(format='yolo'))

#ex1)
transform = A.Compose([
    A.RandomCrop(width=256, height=256, p=1),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
], bbox_params=A.BboxParams(format='yolo', min_area=256))

#ex2)
transform = A.Compose([
    A.RandomCrop(width=256, height=256, p=1),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
],bbox_params=A.BboxParams(format='yolo', min_visibility=0.25))

bboxes = [
  [0.1, 0.1, 0.2, 0.2],
  [0.3, 0.4, 0.1, 0.2],
    :
]

image = cv2.imread('/path/to/image')

transform = A.Compose([
    A.RandomCrop(width=256, height=256, p=1),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
], bbox_params=A.BboxParams(format='yolo'))

result = transform(imgas=image, bboxes=bboxes)

def clamp(n, smallest, largest):
    return max(smallest, min(n, largest))
bboxes = [[0.01, 0.55 , 1.0,0.94, "car"]]

result = A.BBoxSafeRandomCrop(erosion_rate=0.3)(image=img, bboxes=bboxes)
img2 = result['image'].copy()
for bbox in result['bboxes']:
  x1, y1, x2, y2, label = bbox
  print(x1, y1, x2, y2, label)
  h, w, _ = img2.shape
  x1 = clamp(x1, 0, 1)
  y1 = clamp(y1, 0, 1)
  x2 = clamp(x2, 0, 1)
  y2 = clamp(y2, 0, 1)
  cv2.rectangle(img2, (int(x1*w), int(y1*h)), (int(x2*w), int(y2*h)), (0,0,255), 4)

cv2_imshow(img2)