MSER and SIMSER Regions: A Link Between Local Features and Image Segmentation

In this paper, the concept of using salient regions (MSER and SIMSER features) for image segmentation is revised and evaluated. Although we focus on the foreground-background segmentation (which plays an important role of many machine vision problems) the presented results and conclusions are also applicable to more general tasks of segmentation. It is shown that standard MSER features do not provide satisfactory performances in typical segmentation problems, while SIMSER features (which are fully scale-invariant modifications of MSERs) are a more promising tool, with only marginally higher computational costs than MSERs. The presented conclusions are illustrated by exemplary results on a challenging benchmark dataset.