pashaalex/mesh_augmentator

Allows you to bend an image in 3D space. It is also acceptable to change lighting, shadows and depth of field

This project is used to augment paper images based on position, curl, and lighting.
<- Demo without depth of field.

Quick start for augment images

Prepare data for augment

import cv2
import numpy as np
from mesh_augmentator import get_random_augment
from os import listdir
from os.path import isfile, join

# get any control points to be set on image
points_src = []
with open("rects.txt") as file:    
    for line in [line.rstrip() for line in file]:
        x, y, w, h = [float(a) for a in line.split(':')[1].split(';')]
        points_src.append((x, y))

In case of many images with background

image = cv2.imread('sample.jpg')
root = 'bgs'
background_fnames = [join(root, f) for f in listdir(root) if isfile(join(root, f))]

# get result image and control points
image, points = get_random_augment(image, 224, 200, points = points_src, backgrounds = None, background_fnames = background_fnames)

# draw control points on result augmented image
for x, y in points:
    cv2.circle(image, (int(x), int(y)), radius=3, color=(0, 0, 255), thickness=-1)
cv2.imwrite('output1.jpg', image)

In case of one or more background images in memory

image = cv2.imread('sample.jpg')
blank_image = np.zeros((224, 200, 3), np.uint8)

# get result image and control points
image, points = get_random_augment(image, 224, 200, points = points_src, backgrounds = [blank_image], background_fnames = None)

# draw control points on result augmented image
for x, y in points:
    cv2.circle(image, (int(x), int(y)), radius=3, color=(0, 0, 255), thickness=-1)
cv2.imwrite('output2.jpg', image)

Information about mesh render

import cv2
from mesh_augmentator import MeshModel

output_dim = 224
#prepare data
sample = cv2.imread('sample.jpg')
h, w, dc = sample.shape
k = min(output_dim / h, output_dim / w) * 2.2
sample = cv2.resize(sample, (int(w * k), int(h * k)), interpolation = cv2.INTER_LINEAR)
h, w, dc = sample.shape
background = cv2.imread('wood.jpg')
background = cv2.resize(background, (output_dim, output_dim), interpolation = cv2.INTER_LINEAR)

#render
mesh = MeshModel(100, 100, sample, True, True)
mesh.cylynder_vertical(R = w * 8)
mesh.shift(0, 0, mesh.get_best_object_distance())

mesh.set_output_size(output_dim, output_dim)
output_image = mesh.render(background)
cv2.imwrite("output.jpg", output_image)

Depth of field:

Light position change:

Light diameter change:

Shadow:

Rotate and curl: