A shape model with coactivation networks for recognition and segmentation

Abstract

Research in image understanding is an ongoing endeavor in a variety of disciplines and manifests itself in a variety of application domains. The typical goal of this research is to develop techniques to automatically extract meaningful information from a population of images. Object recognition is a key process in image understanding because it enables additional higher-level processing. We present a components-based object detection and localization algorithm that aids segmentation for static images, along with a new approach for the discrimination of object presence in the static domain. The approach integrates the learned knowledge about the category with the supporting information in the image to combine the segmentation and recognition processes, allowing their interactions to guide subsequent processing. An activation network provides topdown perceptual grouping to supplement the initial holistic interpretations associated with the initial hypotheses. Localization is aided by allowing an activated interpretation to trigger an investigation of reinforced interpretations that can support or suppress a current hypothesis, reducing the reliance on low-level feature detectors to achieve proper detection.