The topic concerns image classification: Kaggle APTOS 2019 Blindness Detection competition.
My goal is to crop images to 90% of the image size and have ‘neutral’ padding > no repetition of border areas.
Before using transformations, all images are resized to size 224.
Then I use the following code:
my_xtra_tfms = [flip_lr(p=0.5),
rotate(degrees=(0,360),p=1),
skew(direction=(1,7), magnitude=0.15),
crop_pad(size=200, padding_mode = 'zeros', p=1),
]
tfms = get_transforms(do_flip=True, flip_vert=False, max_zoom=1.2, max_warp = 0.05,
p_affine = 0.25, xtra_tfms=my_xtra_tfms)
data = (src.transform(tfms, size=64)
.databunch(bs=32)
.normalize(imagenet_stats))
The problem is that the resulting images appear to somehow still mirror the boundaries. Further I don’t see the effect of experimenting with different crop sizes! All return images similar to the ones shown:
For the upper left and lower right images, part of the image is repeated.
My goal is to just get this noise out of the image.
If anyone could give some help for better use of crop, size, padding to achieve my goal: thank you!