I-am-Future / GrabCut_reproduce

A reproduce repository to the classic computer graphics paper: "'GrabCut': interactive foreground extraction using iterated graph cuts" (https://doi.org/10.1145/1015706.1015720). Two reproduction versions (paper & OpenCV) are provided.

GrabCut_reproduce

A reproduce repository to the classic computer graphics paper: "'GrabCut': interactive foreground extraction using iterated graph cuts" (https://doi.org/10.1145/1015706.1015720)

In this repo, we have two version of GrabCut: 1 the one fully reproduce from the original paper and 2 the one which took references with OpenCV's GrabCut C++ source code.

• paper version: (grabcut_core_reproduce.py and grabcut_utils_reproduce.py)
  • Reproduction is more similar to the paper, basically followed the equations and algorithms on the paper. It has less performance and speed due to my poor coding ability.
• OpenCV version: (grabcut_core_fromcv2.py and grabcut_utils_fromcv2.py)
  • There are some difference to the original paper. I read the C++ source code and modified it for Python. It has better performance and is very fast. Meanwhile, a UI was designed for the users to perform further fore/background editing.

Demo

The Demo shown below is with the OpenCV modification version.

1

• Brush notation (small area):

• Result:

2

• Brush Notation (large area):

• Result:

Requirements

• Programming Language: python 3.x
• Package Required: numpy, opencv-python (cv2), sklearn, PyMaxflow.

All packages above can be installed with pip install <package name>.

Usage

For two versions of code, some utility functions are in grabcut_utils_xxx.py file, and core part is in grabcut_utils_xxx.py file. In the main.py there is a demo. For the usage, please use fromcv2 version, which is faster (several to tens of seconds), and also has better segmentation quality, and equips with an editing UI to manually set foreground and background after segmentation.