kapadia/geoblend

Geo-aware poisson blending

geoblend's Issues

Image resize leaves null boundaries

When ensuring the source and reference images are of equal pixel resolution, scipy's imresize function sometimes leaves null boundaries.

In [3]: source
Out[3]:
array([[    0, 10322, 10690, ...,  6754,  6562,     0],
       [    0, 10594, 10626, ...,  6481,  6273,     0],
       [    0, 10690, 10770, ...,  6786,  6513,  6594],
       ...,
       [    0,  5137,  5057, ...,  7234,  6978,  7218],
       [    0,  5217,  5265, ...,  7250,  6738,  7394],
       [    0,     0,  5137, ...,  7442,  7490,  7218]], dtype=uint16)

Potentially using scipy's zoom function is better, and avoids calling PIL. Other options is to use imageutil's upsample function.

Documentation

• Get a varied sample set of images for demonstration
• Write out the math

`matrix_from_mask` should support rectangular images

matrix_from_mask creates a coefficient matrix based on a non-rectangular image with bounds designated by a mask.

Due to array length differences, this fails in the case of an mask representing the entire image. Attention to indexing should provide a fix.

In [1]: import numpy as np

In [2]: from geoblend.coefficient_matrix import *

In [3]: mask = np.ones((5, 5))

In [4]: m1 = create_coefficient_matrix(mask.shape)

In [5]: m2 = matrix_from_mask(mask)
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-5-0201c5d24767> in <module>()
----> 1 m2 = matrix_from_mask(mask)

geoblend/geoblend/coefficient_matrix.py in matrix_from_mask(mask)
     73     #
     74     u = np.zeros(n - 1)
---> 75     u[indices] = -1
     76     u[indices[0] - 1] = -1
     77

IndexError: index 24 is out of bounds for axis 1 with size 24

Regularization term

The solution to the poisson equation is second order differentiable. This results in having the blended image (e.g. the solution) be overly smooth in places where discontinuity is desired (e.g. sharp edges around buildings).

A regularization term can help reduce the smoothness based on discontinuities in the gradient.

/cc @dieterichlawson

create_multilevel_solver not defined in cli

After running python setup.py to install geoblend, the create_multilevel_solver is not defined and crashes the cli.
https://github.com/kapadia/geoblend/blob/master/geoblend/scripts/cli.py