Line drawing on empty boolean canvas

[ogencoglu]

Thanks for the interesting work.

How can one replicate the skimage.draw.line and skimage.draw.line_aa line drawing on a boolean empty canvas in zignal? Do you have any performance comparison for this use case for different canvas sizes?

[arrufat]

Hi! First of all, thank you for showing interest in zignal.
While researching this I found (and fixed 77fea30) a bug in the draw_line implementation, so please use zignal 0.4.1 (just released).

I had no idea skimage.draw.line had such a feature, it's very versatile and cool!

There's no direct support in zignal (the Python bindings only allow you to work with RGBA images, because they can easily be optimized with SIMD operations, as 4 u8 can be seen as just a u32).
However, you can do it with one extra line:

Basic Line Drawing (skimage.draw.line → zignal DrawMode.FAST)

# scikit-image
from skimage import draw
canvas = np.zeros((100, 100), dtype=bool)
rr, cc = draw.line(10, 10, 90, 90)
canvas[rr, cc] = True

# Zignal equivalent
from zignal import Image, DrawMode
img = Image(100, 100, 0)
canvas = img.canvas()
canvas.draw_line((10, 10), (90, 90), 255, width=1, mode=DrawMode.FAST)
result = img.to_numpy()[:, :, 0] > 0  # Convert to boolean

Antialiased Line Drawing (skimage.draw.line_aa → zignal DrawMode.SOFT)

# scikit-image
from skimage import draw
canvas = np.zeros((100, 100), dtype=float)
rr, cc, val = draw.line_aa(10, 10, 90, 90)
canvas[rr, cc] = val

# Zignal equivalent
from zignal import Image, DrawMode
img = Image(100, 100, 0)
canvas = img.canvas()
canvas.draw_line((10, 10), (90, 90), 255, width=1, mode=DrawMode.SOFT)
result = img.to_numpy()[:, :, 0] / 255.0  # Normalized float values

But how does it compare to skimage?

Visual inspection

I would say pretty close

Performance comparison

I drew 100 random lines

Fair comparison: including the numpy conversion

Canvas: 100x100
  zignal FAST:  0.1 ms
  zignal SOFT:  0.2 ms
  skimage:      0.5 ms
  skimage_aa:   1.9 ms
  Speedup: 4.9x (FAST), 11.4x (SOFT)

Canvas: 500x500
  zignal FAST:  0.6 ms
  zignal SOFT:  2.6 ms
  skimage:      1.5 ms
  skimage_aa:   7.9 ms
  Speedup: 2.5x (FAST), 3.1x (SOFT)

Canvas: 1000x1000
  zignal FAST:  1.4 ms
  zignal SOFT:  6.7 ms
  skimage:      0.6 ms
  skimage_aa:   12.6 ms
  Speedup: 0.5x (FAST), 1.9x (SOFT)

Raw drawing comparison: just modifying the pixels in the image

Canvas: 100x100
  zignal FAST:  0.1 ms
  zignal SOFT:  0.1 ms
  skimage:      0.5 ms
  skimage_aa:   2.2 ms
  Speedup: 8.0x (FAST), 25.4x (SOFT)

Canvas: 500x500
  zignal FAST:  0.1 ms
  zignal SOFT:  0.4 ms
  skimage:      0.6 ms
  skimage_aa:   7.9 ms
  Speedup: 6.0x (FAST), 17.7x (SOFT)

Canvas: 1000x1000
  zignal FAST:  0.2 ms
  zignal SOFT:  0.9 ms
  skimage:      0.7 ms
  skimage_aa:   18.1 ms
  Speedup: 3.5x (FAST), 20.0x (SOFT)

So, I would say zignal is quite fast, but the headstart is counterbalanced as the image gets larger, since we have to perform more operations to binarize/normalize it in NumPy.

Note that, if you use an RGBA color in the draw_line method, zignal will alpha-blend it with the background automatically. Also, zignal supports arbitrary width lines.

Anyway, I hope this answers your question. It was fun for me to do it :)

[ogencoglu]

Wow, thank you for the very swift reply and awesome analysis. I will now test this out. Appreciated!