Automated annotation scheme for extending bounding box representation to detect ship locations
Havryliuk, Oksana (2023)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
Julkaisun pysyvä osoite on
Bounding boxes often provide limited information about the shape and location of an object on an image. Their limitations lie in their reduced ability to correctly represent objects that have complex shapes or are located at an angle. Related works introduce new object representations that include segmentation masks, keypoints, polylines, and regions and are effective in capturing complex shapes and attributes of an object, but lack computational efficiency for real-time applications and require annotated datasets. The aim of the thesis is to propose an approach to extend bounding box representation to include attributes of interest at a low computational cost. Moreover, the approach aims to automatically transform existing bounding boxes into a new object representation. As a result, the thesis is potentially beneficial for real-time applications that need a complex object representation at a small cost, as well as to create new datasets from existing bounding boxes data or to detect faulty bounding boxes. The approach consists of using a segmentation model to compose a new object representation from a bounding box. The task of object detection is essential in computer vision applications, such as autonomous driving, surveillance, and robotics. The traditional method of representing objects using bounding boxes has limitations in capturing complex shapes and attributes of an object. Therefore, the motivation for this thesis is to propose a low computational cost approach to extend bounding box representation to include attributes of interest. To address this problem, the proposed approach involves using a segmentation model to compose a new object representation from a bounding box. The segmentation model generates a mask for the object, which can be used to extract more detailed features, such as object contours, keypoints, and regions. The approach aims to automatically transform existing bounding boxes into a new object representation, which is potentially beneficial for real-time applications that require a complex object representation at a small cost, as well as to create new datasets from existing bounding box data or to detect faulty bounding boxes. In summary, the approach proposed in this thesis provides an efficient and automated way to improve object representation in computer vision tasks. The experimental results show that the proposed approach achieves better object detection accuracy compared to the traditional bounding box representation, especially for objects with complex shapes and attributes. The approach also has the potential to improve the efficiency of real-time applications that require a complex object representation. Overall, this thesis contributes to the development of more accurate and efficient computer vision systems for various applications.