Bottom-up Object Segmentation for Visual Recognition

Abstract

Automatic recognition and segmentation of objects in images is a central open problem in computer vision. Most previous approaches have pursued either sliding-window object detection or dense classification of overlapping local image patches.<br /> Differently, the framework introduced in this thesis attempts to identify the spatial extent of objects prior to recognition, using bottom-up computational processes and mid-level selection cues. After a set of plausible object hypotheses is identified, a sequential recognition process is executed, based on continuous estimates of the spatial overlap between the image segment hypotheses and each putative class. <br /> The object hypotheses are represented as figure-ground segmentations, and are extracted automatically, without prior knowledge of the properties of individual object classes, by solving a sequence of constrained parametric min-cut problems (CPMC) on a regular image grid. It is show that CPMC significantly outperforms the state of the art for low-level segmentation in the PASCAL VOC 2009 and 2010 datasets. <br /> Results beyond the current state of the art for image classification, object detection and semantic segmentation are also demonstrated in a number of challenging datasets including Caltech-101, ETHZ-Shape as well as PASCAL VOC 2009-11. These results suggest that a greater emphasis on grouping and image organization may be valuable for making progress in high-level tasks such as object recognition and scene understanding.