Waterpixels: superpixels based on the watershed transformation
2013 - 2016
Part of PhD project, under the direction of Etienne Decencière and Thomas Walter
We focus on the partition of the image into small regular homogeneous regions called superpixels. The latter are very useful as they typically serve as primitives for further analysis of the image such as detection, segmentation and classification of objects.
We propose a new method based on the watershed transformation to generate superpixels, called waterpixels.
Click here to see an application of waterpixels to image segmentation learning.
Waterpixels in a nutshell
Waterpixels are based on the marker-controlled watershed (Beucher and Lantuéjoul, 1979; Beucher and Meyer, 1993), a powerful tool for segmentation from Mathematical Morphology. Markers are chosen to improve the pertinence of regions boundaries as well as their spatial regularity (see Fig.c.). The gradient is also improved to encourage regularity (see Fig.e.). Learn more here about waterpixels.