Add fit_output (False/True) option to PerspectiveTransform augmenter

[ZhengRui]

Recalculate the transformation matrix to cover all the input content after transformation.

Sample

fit_output=False:


fit_output=True:

import cv2
import numpy as np
import imgaug as ia


def ia_augmentation_text(images, random_state, parents, hooks):
    for image in images:
        h, w = image.shape[:2]
        y_text = np.random.randint(int(0.75*h), h-20)
        s_text = np.random.randint(20, max(int(0.1*h), 30))
        n_text = 2 * (w-20) // s_text

        image[...] = ia.draw_text(image, y_text, 10, random_string(n_text), color=(255, 255, 255), size=s_text)

    return images

ia_augmenter_text = iaa.Lambda(func_images=ia_augmentation_text)
ia_sometimes = lambda aug: iaa.Sometimes(0.5, aug)
ia_augmenter_img = iaa.Sequential([
    # geometrical
    ia_sometimes(
        iaa.OneOf([
            # iaa.Affine(
            #     rotate=(-30, 30),
            #     shear=(-16, 16),
            #     order=[0, 1],
            #     cval=(0, 255),
            #     mode='constant',
            #     fit_output=True,
            # ),
            iaa.PerspectiveTransform(scale=(0.01, 0.15),
                                     fit_output=True)
        ]),
    ),
    # blur mosaic
    ia_sometimes(
        iaa.OneOf([
            iaa.GaussianBlur((0, 5.0)),
            iaa.MotionBlur(k=15),
            iaa.JpegCompression(compression=(85, 100)),
        ]),
    ),
    # noise blackout
    ia_sometimes(
        iaa.OneOf([
            iaa.AdditiveGaussianNoise(scale=0.2*255, per_channel=0.5),
            iaa.Dropout(p=(0, 0.2), per_channel=0.5),
            iaa.CoarseDropout(0.2, size_percent=0.05, per_channel=0.5),
            iaa.CoarseSaltAndPepper(0.2, size_percent=0.05, per_channel=0.5),
        ])
    ),
    # contrast brightness
    ia_sometimes(
        iaa.OneOf([
            iaa.LinearContrast((0.5, 1.5)),
            iaa.GammaContrast((0.5, 2.0)),
            iaa.AddToSaturation((-50, 50))
        ])
    ),
    # color changes
    ia_sometimes(
        iaa.OneOf([
            iaa.UniformColorQuantization(n_colors=16),
            iaa.Grayscale(alpha=(0.0, 1.0))
        ])
    ),
    # text
    ia_sometimes(
        ia_augmenter_text
    ),
    # horizontal flip
    ia_sometimes(
        iaa.Fliplr(0.5)
    )
])

bbox=np.array([210,163,129,187])
im = cv2.imread('d535880338a040b426d215e35cd122c1.jpg')
cv2.rectangle(im, (bbox[0], bbox[1]), (bbox[0]+bbox[2], bbox[1]+bbox[3]), (0,255,0), 1, cv2.LINE_AA)
images = im[np.newaxis, ...]
bbs=[[ia.BoundingBox(x1=bbox[0], y1=bbox[1], x2=bbox[0]+bbox[2], y2=bbox[1]+bbox[3])]]

while True:
    ims_out, bboxes_out = ia_augmenter_img(images=images, bounding_boxes=bbs)
    im_out = ims_out[0]
    bbox_out = bboxes_out[0][0]
    print(bbox_out)
    cv2.rectangle(im_out, (int(bbox_out.x1), int(bbox_out.y1)), (int(bbox_out.x2), int(bbox_out.y2)), (0,0,255), 1, cv2.LINE_AA)
    cv2.imshow('aug', im_out)
    if cv2.waitKey(0) & 0xFF == ord("q"):
        break

[ZhengRui]

The above example is tested on 0.3.0. When test on current master 9b64c15, the box (red) looks different from before, not sure if it is a recent bug or due to some API change.

[codecov-io]

Codecov Report

Merging #452 into master will decrease coverage by 0.08%.
The diff coverage is 25%.

@@            Coverage Diff             @@
##           master     #452      +/-   ##
==========================================
- Coverage   96.24%   96.16%   -0.08%     
==========================================
  Files          34       34              
  Lines       11080    11092      +12     
==========================================
+ Hits        10663    10666       +3     
- Misses        417      426       +9

[aleju]

Looks good to me. Merged into master. Thanks for the help!

The problem in the second set of images looks like it augments BBs based only on their top-left and bottom-right coordinates instead of using all four corners. I'll fix that.