Adaptels: Scale-adaptive Superpixels
Radhakrishna Achanta, Sabine E. Susstrunk · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 2015
Image segmentation techniques either try to segment an image in a semantically meaningful way, or go to the other extreme by creating small clusters of similar pixels of roughly equal size, called superpixels. The former set of techniques rarely succeeds in "bridging the semantic gap'' while the latter is agnostic to even basic properties like object scale and texture. In this paper we propose the missing intermediate solution to the segmentation problem - how to avoid the over-zealous under-segmentation of traditional algorithms as well as the pessimistic over-segmentation of modern superpixel algorithms. We present a segmentation technique that generates compact clusters of pixels we call Adaptels, which adapt automatically to the local texture and scale of an image. Our algorithm liberates the user from making the difficult choice of the right superpixel size. Despite making no assumptions about the semantics of the image, the resulting segments provide a powerful abstraction of the image that has wide-ranging applications. The algorithm is simple and requires just one input parameter. On segmentation comparison benchmarks it proves to be superior to the state-of-the-art. In addition, it is computationally very efficient, approaching real-time performance, and is easily extensible to three-dimensional image stacks and video volumes.