A fast, scalable, and extensive implementation of the Poisson Image Editing paper by Perez et al. 2003.

• Poisson image editing functionalities
  • Seamless cloning - importing gradients
  • Seamless cloning - mixing gradients
  • Seamless tiling
  • Texture flattening
  • Local illumination change
  • Local color change
• Supports all sparse linear system solvers in scipy.sparse.linalg
• Acceleration of large input problems using multigrid approaches
• An Object-Oriented Programming approach for Poisson image editing - see branch oop for more details
  • All editing functions inherit the abstract PoissonImageEditor class

There are many open source Python implementations of Poisson image editing. However, most implementations only focus on image blending, while ignoring other Poisson image editing applications listed in the paper. This implementation aims to faithfully reproduce all experiments and results presented in the paper. The following table shows the implemented functionalities:

Furthermore, this implementation is signifacantly faster and scales much better than others. The following table shows the profiled run time of seamless cloning on different datasets (in seconds):

The following figure shows the scaling performance of this implementation compared to that of Src. 3. For fair comparison, both implementations are modified to use the same solver scipy.sparse.linalg.spsolve.

By using multigrid solvers, editing a 1080p image can be done in less than 30 seconds, a 6.7x speedup compared to the current fastest open source Python implementation:

To run all experiments using the given datasets (contains testing images in the paper and this README file), run

./run_experiments.sh

To test on your own dataset, run

python3 seamless_cloning.py --help

• data_dir: Folder that contains the input image files. The folder should contain image files named mask, target, and source. The file extension of the files can be arbitrary, as long as the files are valid image files.
• grayscale: Whether to perform blending on the grayscale images.
• solver: Linear solver to use when solving the poisson blending problem. The value of solver should either be function names in the scipy.sparse.linalg library, or "multigrid". Default is spsolve.
• gradient_mixing_mode: Method to mix source and target image gradients. max implements 3. Seamless cloning - Mixing gradients section in the paper, while alpha + gradient_mixing_alpha == 1.0 implements 3. Seamless cloning - Importing gradients section.
• gradient_mixing_alpha: Alpha to blend source and target image gradients. Has an effect only when gradient_mixing_mode == "alpha".

To test on your own dataset, run

python3 seamless_tiling.py --help

To test on your own dataset, run

python3 texture_flattening.py --help

• use_edge: Use edge.* edge map image file in the folder specified in the data_dir folder. If this flag is not set, then computes the edge map from provided source image using Canny edge detector and binary dilation.
• canny_threshold: Thresholding parameters for Canny edge detector. You can play with this parameter for different flattening results. See the documentation for more information.
• edge_dilation_kernel: Kernel size to dilate detected edges. The kernel is a square box filter filled with ones.

To test on your own dataset, run

python3 local_illumination_change.py --help

• data_dir: Folder that contains the input image files. The folder should contain image files named mask and source. The file extension of the files can be arbitrary, as long as the files are valid image files. The illumination of the regions specified by the mask will be modified.

To test on your own dataset, run

python3 local_color_change.py --help

• mode: Color change mode. If mode == "gray_background", then pixels outside the masked region will be converted to grayscale. If mode == "color_change", the hue of the masked region is increased by the value specified by change_hue parameter.
• change_hue: Value added to the hue channel of the masked region.

Source	Target	Mask	Result

Texture	Naive tile	Seamless tile

Source	Mask	Edge	Flattened

Source	Mask	Modified

Source	Green-ish	Blue-ish	Gray background

• Faster solvers: spsolve, cgs, bicg
• minres gives bad results.
• If you want to use conjugate gradient solvers, use bicg, bicgstab or cgs. Do not use solver == "cg" since the A matrix is not hermitian (or symmetric since A is real).
• Iterative least-squares solvers lsqr, lsmr tend to be much slower.