Supervised Multispectral Image Segmentation
Welcome to our dataset for supervised multispectral image segmentation. It comes with ground-truth segmentation, seed point placements and experimental results. This dataset was created for the publication "Supervised Multispectral Image Segmentation With Power Watersheds" by J. Jordan, E. Angelopoulou. ICIP 2012.
Click here to browse the experimental data. Click here for download.
Click here for the segmentation source code.
Citation:
Please cite the following paper when using the data:
J. Jordan and E. Angelopoulou,
"Supervised multispectral image segmentation with power watersheds,"
in Image Processing (ICIP), 2012 19th IEEE International Conference on,
2012, pp. 1585-1589.
Notes on the data:
- The images used in our experiments are from the CAVE Multispectral Image Database.
- As the first four pixel columns of each image do not reflect the scene, but instead appear to consist of random noise, we crop them out with the following command before creating ground-truth segmentations:
for i in */*_ms.png; do mogrify $i -crop 508x512+4+0 +repage; done; - The colors of ground-truth images do not matter.
- In the seed point images, black pixels are foreground seeds, white pixels are background seeds. The different shades of gray are not visible to the algorithms.
Algorithms used to generate segmentations:
Similarity Measure | Algorithm |
L∞ | power watersheds, q=2 |
SA | power watersheds, q=2 |
SID | maximum spanning forest |
SIDSAM1 | maximum spanning forest |
NED | power watersheds, q=2, geod. |
SOM | power watersheds, q=2, geod. |