The goal of the package puppet-warp is provide plug and play solution for image transformation similar to Adobe Photoshop Puppet Warp tool. Since Photoshop solution is proprietary, hence any scripting might be a big issues especially in environments where Photoshop is not supported, I decided to create this package based on Python in which Puppet Warp is programmatically manageable and used in automation processes where advanced transformation method is required.

• As-Rigid-as-Possible Shape Manipulation of triangular mesh
• Image transfer from triangualar mesh at rest to mesh defined by ARAP transformation

numpy>=1.21.5
opencv-contrib-python>=4.5.4.60,<=4.7.0.72
opencv-python>=4.5.4.60,<=4.7.0.72
scikit-image>=0.19.2,<=0.21.0
scikit-learn>=1.0.2,<=1.2.2

Optional:

triangle>=20200424

pip install puppet-warp

For latest version from git, I recommend you to use:

pip3 install git+https://github.com/mikecokina/puppet-warp.git@master

Install with Jonathan Richard Shewchuk's (Dzhelil Rufat) triangle:

pip install puppet-warp[jrs]

The package puppet-warp is coming with live demo. To open lived demo, use following code.

from pwarp import Demo
Demo().run()

To be able manipulate with image, select control points by clicking on vertices in image (connections of edges). When points are chosen, take single points by cursor and drag the point wherever you like.

Demo comes with capability to store transformed mesh. To save the mesh, hit Space Bar. To quit demo, hit Esc button.

import cv2
from pwarp import triangular_mesh, Demo
from pwarp._io import save_wavefront

# DEFINE WITHD and HEIGHT of your example image and DELTA step to create triangular mesh.
width = 800
height = 492
delta = 100
method = 'scipy'  # or jrs

# DEFINE path to your image and path, where WAVEFRONT file will be stored.
wavefront_path = "image.obj"
image_path = "image.jpg"

# noinspection PyUnresolvedReferences
image = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB)

# This will genearete triangular mesh over the image with given
r, f = triangular_mesh(width=width, height=height, delta=delta, method=method)
# Save wavefront object.
save_wavefront(wavefront_path, no_vertices=len(r), no_faces=len(f), vertices=r, faces=f)

Demo(
    image=image_path,
    obj_path=wavefront_path,
    screen_height=height,
    screen_width=width,
    scale=1,
    dx=0,
    dy=0,
    verobse=True
).run()

The graph warp requires vertices and faces (triangulation), control points and new displacement of control points. Based on the given information, graph transform compute new positions of supplied vertices.

Example:

from pwarp import get_default_puppet, graph_warp

# Control points represent indices of points in original vertex array.
control_pts = np.array([22, 50, 94, 106], dtype=int)
# Shif represents new positions of control points respectively to `control_pts` list.
shift = np.array(
    [[0.555, -0.905],
     [-0.965, -0.875],
     [-0.950, 0.460],
     [0.705, 0.285]], dtype=float
)
puppet = get_default_puppet()
new_vertices = graph_warp(
    vertices=puppet.r,
    faces=puppet.f,
    control_indices=control_pts,
    shifted_locations=shift
)

The graph defined warp will transform areas of image covered by source vertices to given destination vertices. An algorithm requires image, source and destination vertices, and faces for both. An order of faces (triangles) in both sets have to be same, so in other words, source and destination faces must form pairs. A pixel in each triangle is transformed via affine transformation defined by source to destination face.

Example:

import cv2
from matplotlib import pyplot as plt
from pwarp import graph_defined_warp, graph_warp, get_default_puppet

control_pts = np.array([22, 50, 94, 106], dtype=int)
shift = np.array(
    [[0.555, - 0.905],
     [-0.965, - 0.875],
     [-0.950, 0.460],
     [0.705, 0.285]], dtype=float
)
puppet = get_default_puppet()
new_r = graph_warp(
    vertices=puppet.r,
    faces=puppet.f,
    control_indices=control_pts,
    shifted_locations=shift
)

image = cv2.cvtColor(cv2.imread("../data/puppet.png"), cv2.COLOR_BGR2RGB)
width, height = 1280, 800
dx, dy = int(width // 2), int(height // 2)
scale_x, scale_y = 200, -200
r = puppet.r.copy()
r[:, 0] = r[:, 0] * scale_x + dx
r[:, 1] = r[:, 1] * scale_y + dy

new_r[:, 0] = new_r[:, 0] * scale_x + dx
new_r[:, 1] = new_r[:, 1] * scale_y + dy

image_t = graph_defined_warp(
    image,
    vertices_src=r,
    faces_src=puppet.f,
    vertices_dst=new_r,
    faces_dst=puppet.f
)

fig, axs = plt.subplots(1, 2, frameon=False)
plt.tight_layout(pad=0)

axs[0].imshow(image)
axs[1].imshow(image_t)
axs[0].triplot(r.T[0], r.T[1], puppet.f, lw=0.5)
axs[1].triplot(new_r.T[0], new_r.T[1], puppet.f, lw=0.5)

for ax in axs:
    ax.set_xlim([380, 900])
    ax.set_ylim([150, 750])
    ax.invert_yaxis()
    ax.axis('off')
plt.show()

The algorithm is intended to generate a triangular mesh within rectangle defined by its width and height. The density of the mesh is adjustable via delta parameter. Algorithms is based on generation of frame. Frame is defined by vertices where distance between each two vertices is defined by mentioned delta parameter. The area of frame generated in such manner is triangulated.

Following example will generate mesh within rectangle of dimensions W x H = 1280 x 800 pixels.

Example:

from pwarp import triangular_mesh
r, f = triangular_mesh(width=1280, height=800, delta=100)

Example on full screen triangular mesh warp:

[1] https://www-ui.is.s.u-tokyo.ac.jp/~takeo/papers/takeo_jgt09_arapFlattening.pdf
[2] https://github.com/deliagander/ARAPShapeManipulation.git
[3] https://learnopencv.com/warp-one-triangle-to-another-using-opencv-c-python/
[4] https://rufat.be/triangle/
[5] http://www.cs.cmu.edu/~quake/triangle.html

@article{journals/jgtools/IgarashiI09,
    author = {Igarashi, Takeo and Igarashi, Yuki},
    ee = {http://dx.doi.org/10.1080/2151237X.2009.10129273},
    journal = {J. Graphics, GPU, & Game Tools},
    number = 1,
    pages = {17-30},
    title = {Implementing As-Rigid-As-Possible Shape Manipulation and Surface Flattening.},
    url = {http://dblp.uni-trier.de/db/journals/jgtools/jgtools14.html#IgarashiI09},
    volume = 14,
    year = 2009
}

or

@article{10.1145/1073204.1073323,
    author = {Igarashi, Takeo and Moscovich, Tomer and Hughes, John F.},
    title = {As-Rigid-as-Possible Shape Manipulation},
    year = {2005},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    volume = {24},
    number = {3},
    doi = {10.1145/1073204.1073323},
    journal = {ACM Trans. Graph.},
    month = {jul},
    pages = {1134–1141},
    numpages = {8}
}