Edgeflow-driven Variational Image Segmentation: Theory and Performance Evaluation

1) Paper (Submitted for publication): See this Technical Report

2) Complete data (segmentations) for the evaluation of 6 variational segmentation methods:

• Download complete results on test and training sets (about 13000 segmentation generated in 2 months running on 5 to 10 CPUs).

This package contains about 13000 segmentations run on test
and training images from Berkeley Segmentation Data Set (BSDS).

- Segmentations results and their evaluations (precision, recall, f-measure)
for all methods on both training and test images can be found under the
following directories:
"curveevolution_grayscale", "anisodiffusion_grayscale", "anisodiffusion_color"

- The alpha values (for curve evolutions), gamma values (for anisotropic diffusion)
and value of K (for Perona-Malik flow) are embedded into the file names of the .gif files.
alpha and gamma values are multiplied by 10 to avoid fractions as much as possible.
You need to divide these values by 10 to extract the real alpha and gamma.

- Html pages for side by side visualization of segmentations on test images are provided under the directory:
"html_for_visualization"
(perl file--htmlgenerator.pl--that generated these html pages is also provided).

- The directories "test_images" and "training_images" contain the original images used in these evaluations.

• Browse results on the test set (100 images) of Berkeley Segmentation Data Set.

3) Evaluation Software: Matlab code that is used to optimize segmentation parameters and evaluate segmentations will be made available here soon (This code is partially based on the code provided with Berkeley Segmentation Data Set).

4) Segmentation Software: Results in this paper are generated by a software written in C#. This software is not being made available because of the packaging issues (Contact me if you need access to it).

A new version of the software (segmentation results may be different) written in C++ is available for download.