|
This page allows you to download data associated with our two papers on figure-ground assignment
in natural images. An archive containing the figure-ground labellings is available [fgdata.tar.gz]
along with additional MATLAB code for manipulating the files [fgcode.tar.gz].
Each figure-ground labeling is associated with a particular segmentation
from the Berkeley
Segmentation DataSet (BSDS). You will also likely want to download the set of BSDS
[images],
[segmentations], and the accompanying [code].
The figure-ground labels came from a separate set of human subjects from those
that originally segmented the BSDS images. The sub-directories in the
figure-ground data indicate the BSDS user id, the filename gives the image id
followed by the figure/ground user id. There are two f/g labellings per
segmentation.
To load a figure-ground labeling into MATLAB one may proceed with:
seg = readSeg('/home/eecs/fowlkes/BSDS300/human/color/1103/181091.seg');
bmap = seg2bmap(seg);
[cmap,cid2sid] = seg2cmap(seg,bmap);
FG = readbin('/home/eecs/fowlkes/fgdata/color/1103/181091-104.fg');
This example loads image #181091 which has been segmented by subject #1103 and
each contour subsequently labeled by figure-ground subject #104. FG{1} gives
the figure/ground labeling encoded with two-pixel wide contours. A alternate
representation is contained in FG{2} which gives the label for each contour,
the support of the ith contour is given by (cmap==i) and the pair of abutting
segments by cid2sid.
If you make use of this data, please cite the following paper:
-
C. Fowlkes, D. Martin, J. Malik.
"Local Figure/Ground Cues are Valid for Natural Images"
Journal of Vision, 7(8):2, 1-9.
[pdf]
You may also be interested in:
-
X. Ren, C. Fowlkes, J. Malik.
"Figure/Ground Assignment in Natural Images",
ECCV, Graz, Austria, (May 2006).
[pdf]
|