Different general types of images:

  • Blurry numbers on solid color/textured backgrounds (number taking up most of image) or in a shape in the image
  • Small blurry numbers on architectural backgrounds (on framing, in windows, on garage doors) where number is really tiny in the image
  • Numbers with decoration around them (like placards with the house number)
  • Numbers partially hiding behind plants

Script to get heights/widths of all images:

import glob
from PIL import Image

for filepath in glob.iglob('train/*.png'):
	im=Image.open(filepath)
	print(im.size[0]) # 0 is width, 1 is height

and then run:

python3 stats.py > stats.txt

Isolations:

  • Geometric shapes
  • Architectural patterns (brick, stone, etc)
  • Different types of number forms (handwriting, different fonts, etc)
  • Placards and other types of signs the numbers are on

Modifications (global):

  • 3D transform
  • Motion blur
  • Brightness/contrast

Modifications (local):

  • Shadows