Data Preparation

[Ritsu-mio]

I faced this problem When use the code reference_video. How do I prepare the distance field images of lines arts?

[lento]

I'm sorry for the inconvenience.

I have forgotten to explain how to get distance field images of line arts. In order to obtain them, I have used snowy like below. Later, I will upload the code and fix the documents.

import numpy as np
import imageio
import snowy


def distance_field(path):
    img = imageio.imread(path, 'JPG-PIL', pilmode='RGBA')
    img = np.clip(np.float64(img) / 255, 0, None)
    mask = np.zeros((img.shape[0], img.shape[1], 4))
    mask[img < 0.95] = 1.0
    img = np.expand_dims(mask[:, :, 0], axis=2)
    sdf = snowy.unitize(snowy.generate_sdf(img != 0.0))

    return sdf

# input_path: line art path
# out_filename: output path

dist_image = distance_field(str(input_path))
snowy.export(dist_image, str(out_filename))

[Ritsu-mio]

That will help me a lot. Thank you

[Ritsu-mio]

I am sorry I have the another problem.
Does this apply to get distance field images of grayscale images as well?

[lento]

The code above applied to .png image file. The shape of array gotten by cv2.imread(png_file_path) was (height, width, 3).

Judging from API Reference of generate_sdf, generate_udf function that is used in generate_sdf takes a grayscale image as the input like below. So you may be able to get distance field images without converting RGBA into grayscale like I did in the code above.

def generate_sdf(image: np.ndarray, wrapx=False, wrapy=False):
    a = generate_udf(image, wrapx, wrapy)
    b = generate_udf(image == 0.0, wrapx, wrapy)
    return a - b

def generate_udf(image: np.ndarray, wrapx=False, wrapy=False):
    assert image.dtype == 'bool', 'Pixel values must be boolean'
    assert len(image.shape) == 3, 'Shape is not rows x cols x channels'
    assert image.shape[2] == 1, 'Image must be grayscale'
    return _generate_edt(image, wrapx, wrapy)

[Ritsu-mio]

thank you for your reply.