difficult segmentation data code

This webpage provides the images used in the segmentation experiments in [1]. These segmentations are particularly difficult for graph cuts, as the boundary of the true image corresponds to a cut with large weight, either because the object has long, fine segments (i.e., a long boundary), or shading (i.e., large edge weights across the boundary).
For details, see [1], where these problems are remedied by coupling edges inthe graph cut, and granting selective disounts.

The data below is free to use. If you use the data, please cite the article [1]. The data consists of jpg or png images. For each image, there is the image itself, a user labeling, and a ground truth segmentation (hand-segmented in Adobe Photoshop CS4/CS5). There are three parts:
  1. Grayscale images with shading: objects in grayscale images that are partially shaded;
  2. Color images with shading: objects in color images with partial shading;
  3. Color images for shrinking bias: color images of objects with fine structures.

[1] S. Jegelka and J. Bilmes. Submodularity beyond submodular energies: coupling edges in graph cuts. CVPR 2011. (here is the supplementary material)

(The code for image segmentation with cooperative cuts is available here).


Shaded grayscale images



download images
 download labels
 download "ground truth"



Shaded color images



download images
 download labels
 download "ground truth"



  Objects with fine structures



download images
 download labels
 download "ground truth"


Data instructions

Each of the links above leads to a .zip file with the images, user labels or ground truth of a data set. The corresponding files are all named identically, so they should be extracted into different directories. As an example, the vacuum cleaner would be '026_vacuum.png' in all three zip files. The ground truth labels for computing the twig error are in files ending in 'L' (only for the fine color objects), e.g., '026_vacuumL.png'.
In the grayscale set, user labels are as follows: white (255) is foreground, gray (125) is background, and everything else is unlabeled. For the color images, the labeling is: red is foreground, blue is background, and everything else (white) is unlabeled. The "ground truth" images also have some grayscale labeled pixels -- for those, the labeling is not so clear, and we ignored those pixels in the experiments ('object' was pixels with value > 229, 'background' was pixels with value < 26).
Here is a list of the images.