jianchaotan / decompose-single-image-into-layers

Decomposing Images into Layers via RGB-space Geometry

This code implements the pipeline described in the paper "Decomposing Images into Layers via RGB-space Geometry" by Jianchao Tan, Jyh-Ming Lien, and Yotam Gingold in ACM Transactions on Graphics (TOG).

The pipeline is divided into two steps.

1. Convex Hull Simplification

Input:

• Source image

Output:

• Simplified Hull Vertices (color palettes) for multiple simplification levels (4 to 10 vertices)

Users can then choose what simplification level (number of vertices) is reasonable. They can visualize the output by dragging-and-dropping it onto our web GUI.

2. Layer Extraction

Input:

• Source image
• Vertex order (users choose which they like)
• Simplified convex hull vertices (the color palettes)
• Optimization weights

Output:

• Translucent layers (PNG's)
• Barycentric Coordinate weights (a JSON file and also PNG's of each weight map for visualization)

Users can perform global recoloring in our web GUI. First load the original image, then drag-and-drop the convex hull .js file, and finally drag-and-drop the Barycentric Coordinates weights .js file. For a more detailed usage guide to the web GUI, please see the supplemental materials of our paper.

We provide two choices for this layer extraction step:

• Our main global optimization-based approach, including RGB and RGBA versions. The RGB optimization assumes that the first layer is opaque, which means its color is the background color. The RGBA version is appropriate when input image is RGBA translucent image or there is no obvious background color in input image, such as a photograph.
• Our As-Sparse-As-Possible (ASAP) approach, which is independent per-pixel and therefore much faster but generates noisier results.

Example usage

Convex Hull Simplification:

cd examples
python ../ConvexHull_Simplification/SILD_convexhull_simplification.py apple

Layer extraction (RGB):

cd examples
python ../Layer_Extraction/SILD_RGB.py apple.png apple-06-vertex_order.js apple-final_simplified_hull_clip-06.js apple-06-layers-RGB --weights weights.js

Layer extraction (RGBA): (Note that apple has an opaque background, so it is not a good example for RGBA layer extraction.)

cd examples
python ../Layer_Extraction/SILD_RGBA.py moth.png moth-06-vertex_order.js moth-final_simplified_hull_clip-06.js moth-06-layers-RGBA --weights weights.js

Layer extraction (ASAP) (all arguments are in the JSON parameter file):

cd examples
python ../Layer_Extraction/SILD_ASAP.py apple-06-ASAP.js

You can test if your installation is working by comparing your output to the apple-output directory. You can find the output for other examples in the supplemental materials of our paper. If you want the optimization to be faster, you can increase gtol and ftol values in scipy.optimize.minimize, which is like: options={'ftol':1e-4, 'gtol': 1e-4}. Our current implementation in paper is using default values of the function, so it is sometimes slow.

The apple.png image is copyright Adelle Chudleigh. The moth.png image is copyright Michelle Lee.

Dependencies

• NumPy
• SciPy
• Cython
• GLPK (brew install glpk)
• cvxopt, built with the GLPK linear programming solver interface (CVXOPT_BUILD_GLPK=1 pip install cvxopt)
• PIL or Pillow (Python Image Library) (pip install Pillow)