Author:

Keywords:

PSI_VISICS, PSI_4065

Abstract:

Image segmentation has mostly been approached in a bottom-up fashion using low-level cues such as color, texture or motion. More recently, top-down segmentation has been explored as well, solving the segmentation task in combination with object detection. Here, we want to extend these approaches from the individual object-level to the scene-level. A scene is typically composed of multiple objects, both things and stuff, that possibly occlude one another. Rather than running several object detectors/segmentors in parallel and independent of one another, the goal of this work is to develop a global optimization, where a pixel can only be assigned to a single object and the aim is toexplain the whole image.