Unsupervised Color Segmentation

You can find more details (including examples, pre-calculated Berkeley results and code) on our team homepage at Virtual Habitat: http://vh.icg.tugraz.at

The problem of segmentation is to partition an image into a set of non-overlapping regions and is one of the most investigated areas in computer vision. Traditionally, segmentation is formulated as bottom-up process, where no high-level knowledge about the image scene is incorporated into the algorithm. Bottom-up approaches identify regions in the input image only based on low-level cues, like color or texture. Although such segmentation results can be achieved in an efficient way, they often do not match manual segmentations. To overcome the limitations of low-level cues based approaches, there has recently been much interest on top down algorithms or on simultaneous combination of top-down and bottom-up approaches. But there are still many applications where a priori information is hard to obtain or is not available at all and thus, have to rely on an efficient and as good as possible bottom-up segmentation.

We have introduced two different unsupervised color segmentation concepts (ROI-SEG and SAL-SEG), which are both based on the underlying idea of combining a set of differently focused sub-segmentations into the final result. The two methods mainly differ in the way the sub-segmentations are found. ROI-SEG use Maximally Stable Extremal Region results for efficiently segmenting images, whereas SAL-SEG uses global optimal total variation segmentation to provide even more accurate results. Due to the general design of the approaches, an extension to 3D data sets is also straightforward.

A comprehensive experimental evaluation on the Berkeley image database showed that state-of-the-art results are achieved at reduced computational costs. Results for Probabilistic Rand Index (PRI), Variation of Information (VoI), Global Consistency Error (GCE) and Boundary Displacement Error (BDE) are shown, where the best two results are always highlighted in bold.