
As an Amazon Associate and affiliate partner, Menrva Books earns from qualifying purchases. Learn more
Locality is taken as the surrounding or nearby region of something in the world. Without the notion of spatial order that it brings, the world cannot be understood. This thesis is dedicated to exploring locality present in the world, and specifically the visual world, as captured by digital images and videos for the purpose of designing better visual recognition algorithms. We consider traditional computer vision challenges like image classification and segmentation, attribute classification, action recognition and object tracking, which are not fully aware of locality, and ask ourselves the question: How can visual recognition profit from locality? We start by decomposing objects locally for efficient image classification and segmentation, and continue with attribute classification by exploiting locality. Then we propose a video representation considering both spatial and temporal locality in a video for action recognition, and end by enriching object tracking by locality specified with natural language description. Findings of this thesis may lead to a better understanding of how to explore locality for a wide variety of other challenges in computer vision as well.--Samenvatting auteur.
This thesis investigates how the concept of locality, or the spatial ordering of elements, can be leveraged to improve visual recognition algorithms. The author explores the application of locality in various computer vision tasks, including image classification, segmentation, attribute classification, action recognition, and object tracking. By decomposing objects locally and considering spatial-temporal locality in videos, the research aims to enhance the performance of algorithms that are not fully locality-aware. The findings suggest potential improvements for a wide range of computer vision challenges.
This work presents a focused exploration of locality within the field of computer vision, aiming to enhance existing algorithms. The thesis systematically addresses several core computer vision challenges, proposing novel approaches that integrate the concept of spatial and temporal locality. The research is presented as a foundational study, with findings intended to inform future developments in visual recognition systems.
Page Count:
99
Publication Date:
2017-01-01
ISBN-10:
9461828543
ISBN-13:
9789461828545
No comments yet. Be the first to share your thoughts!