Implementation of the Waterpixels algorithm from
Vaïa Machairas, Matthieu Faessel, David Cárdenas-Peña, Théodore Chabardes, Thomas Walter, et al.. Waterpixels. IEEE Transactions on Image Processing, Institute of Electrical and Electronics Engineers, 2015, 24 (11), pp.3707 - 3716. 10.1109/TIP.2015.2451011 . hal-01212760
To compile, run the following lines on a terminal window, from the root folder:
mkdir build
cd build
cmake ..
make
cd ..
If you want to get metrics for the resulting segmentations of a dataset, create a directory named images and one named human-labels, both children of the root folder. Place the images straight inside the images folder. Then, place the .seg segmentation files inside the human-labels folder, but separate them into folders for each user. For example:
human-labels/1102/33039.seg
That's the segmentation file for image 33039.jpg made by user 1102.
To display the superpixel segmentation result from a single image, run the following line:
./build/waterpixels_example <file name> <superpixel count>
For example:
./build/waterpixels_example 302003.jpg 100
To get all boundary adherence metrics from a dataset, run
./build/waterpixels_metrics adherence
This will save the ground-truth segmented image under the folder results.
For example:
results/1102/33039_gt.png
This will also save the distribution of Manhattan Distances from each ground truth border pixel to the closest superpixel segmentation border pixel. For example, for a superpixel count of 50:
results/1102/33039_dist50.txt
And finally, this will also save superpixel segmentation under results/segmentation.
To get the contour density and average mismatch factor of the resulting segmentations, simply run:
./build/waterpixels_metrics regularity
This will save these metrics under results/measures, in a .csv format.
To get the time it takes to compute the segmentations of the entire dataset, run:
./build/waterpixels_metrics time
The result will be saved at results/measures/time.csv