Contour-Based Object Extraction and Clutter Removal for Semantic Vision

Abstract

This paper focuses on object extraction from images, a functionality that can be relevant both for category learning and object recognition in diverse applications. The described object extraction approach, which doesn’t take into account any prior knowledge about the target objects, works on the edge-based counterpart of the original image. In a first step, groups of neighboring edge pixels are traced to form contour segments. These contour segments are then coherently aggregated to reconstruct the shapes of the different objects present in the original image. The approach is particularly relevant for extracting objects with few if any distinctive local features, thus objects mainly characterized by their shape. The developed functionalities can be used to segment and extract objects from images with multiple objects, as those obtained from the Internet by searching for a specific object category name. They can also be used to discard clutter from image sub-windows expected to contain a single object, as those delivered by an object detector.