Affine augmentations

Old notes on affine augmentations. Enjoy!

Affine Augmentations

Most of the standard Geometric 2D augmentations can be composed from affine matrices.

Lets load an image with bbox and keypoint labels

import cv2
import yaml
from PIL import Image

# Read image
img = cv2.imread('./augmentation_data.png')
print(f"Image size = width:{img.shape[1]} pixels / height:{img.shape[0]} pixels")

# Read labels
labels = yaml.safe_load(open('./augmentation_data_labels.yml'))['labels']
print(labels)

Image size = width:599 pixels / height:464 pixels
{'wickets': {'x1': 253, 'y1': 250, 'x2': 315, 'y2': 434}, 'bat': {'x1': 372, 'y1': 18, 'x2': 482, 'y2': 69}, 'wicket_camera': {'x': 288, 'y': 386}, 'r_eye': {'x': 314, 'y': 108}, 'l_eye': {'x': 329, 'y': 103}}

Plot labels

#Wickets
image = cv2.rectangle(img, (labels['wickets']['x1'], labels['wickets']['y1']),\
    (labels['wickets']['x2'], labels['wickets']['y2']), (0, 0 , 255), 2)
#Bat
image = cv2.rectangle(img, (labels['bat']['x1'], labels['bat']['y1']),\
    (labels['bat']['x2'], labels['bat']['y2']), (0, 255 , 255), 2)
# Wicket camera
image = cv2.circle(img, (labels['wicket_camera']['x'], labels['wicket_camera']['y']), \
    2, (0, 255, 0), 1)
# l_eye camera
image = cv2.circle(img, (labels['l_eye']['x'], labels['l_eye']['y']), \
    2, (0, 255, 0), 1)
# r_eye camera
image = cv2.circle(img, (labels['r_eye']['x'], labels['r_eye']['y']), \
    2, (0, 255, 0), 1)

pil_image = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
display(pil_image)

Affine Matrix Calculator class

This class composes a sequence of rotation, translation, scaling and shearing into one matrix.

from typing import Any
import numpy as np

class AffineMatrixCalculator:

    def __init__(self):
        pass

    def __call__(self, rotation=0., translation=(0., 0.) ,scale=(1., 1.), shear=(0., 0.)):
        self.M_R = self.getRotationMatrix(rotation)
        self.M_t = self.getTranslationMatrix(translation)
        self.M_Sc = self.getScalingMatrix(scale)
        self.M_Sh = self.getShearingMatrix(shear)
        return self.M_t @ self.M_R @ self.M_Sc @ self.M_Sh

    @staticmethod
    def getRotationMatrix(deg):
        angle = np.radians(deg)
        R=np.array([[np.cos(angle), -np.sin(angle), 0],\
                    [np.sin(angle), np.cos(angle), 0],\
                    [0, 0, 1]])
        return R

    @staticmethod
    def getTranslationMatrix(trans):
        M = np.eye(3)
        M[0, 2] = trans[0]
        M[1, 2] = trans[1]
        return M

    @staticmethod
    def getScalingMatrix(s):
        M = np.eye(3)
        M[0, 0] = s[0]
        M[1, 1] = s[1]
        return M

    @staticmethod
    def getShearingMatrix(s):
        M = np.eye(3)
        M[0, 1] = s[0]
        M[1, 0] = s[1]
        return M

A = AffineMatrixCalculator()
M = A(rotation=10., translation=(0.1, 0.2), scale=(0.7, 0.9), shear=(0.02, 0.05))
print(M)

[[ 0.68155126 -0.14249605  0.1       ]
 [ 0.16587007  0.88875805  0.2       ]
 [ 0.          0.          1.        ]]

Warping the image with the M matrix

aug_img = cv2.warpAffine(img, M[:2, :], (img.shape[1],img.shape[0]))
pil_image = Image.fromarray(cv2.cvtColor(aug_img, cv2.COLOR_BGR2RGB))
display(pil_image)

Warping the labels

aug_labels = {}
w = int(abs(labels['wickets']['x1'] - labels['wickets']['x2']))
h = int(abs(labels['wickets']['y1'] - labels['wickets']['y2']))
wickets = np.array([[labels['wickets']['x1'], labels['wickets']['y1'], 1.], \
                    [labels['wickets']['x1']+w, labels['wickets']['y1'], 1.], \
                    [labels['wickets']['x1']+w, labels['wickets']['y1']+h, 1.], \
                    [labels['wickets']['x1'], labels['wickets']['y1']+h, 1.]])

# transforming points
augmented_wickets = (M @ wickets.T).T

# plotting
aug_img = cv2.polylines(aug_img, pts=np.int32([augmented_wickets[:, :2]]), isClosed=True, color=(0, 0, 255), thickness=2)
pil_image = Image.fromarray(cv2.cvtColor(aug_img, cv2.COLOR_BGR2RGB))
display(pil_image)