Object localization algorithms aim at finding out what objects exist in an image and where each object is. Object localization is fundamental to many computer vision problems. Simply knowing what is in the image is not enough when we want to reason about object properties such as shape, color or pose, and infer the relationship between different objects. Object localization is also a challenging task. Intra-class variance as well as differences in illumination, viewing distance and viewpoint all make objects within the same category look different from each other.The goal of this thesis is to study the problem of object localization in natural images. We start off by answering the three key questions about object localization: "how" good are our algorithms at localizing objects, "what" are the types of mistakes our algorithms commonly make, and "why" are certain objects hard to detect by providing a set of annotations and tools to facilitate analysis on object localization algorithms. One of the conclusions drawn from our analysis is that object detectors sometimes succeed at finding the rough location of the object but fail to precisely locate the object. Inspired by this observation, we proposed a framework that takes as input a roughly localized object bounding box, and generates a more accurate localization of the object, both in the form of a bounding box and pixel-wise segmentation. Finally, we try to explore an alternative way to localize objects, by detecting object boundaries, which could generalize to localizing objects of unknown categories.
PDF
download
878987797
028-85220240
