E-Locus - Institutional Repository of the University of Crete

Home Collections Type of Work Post-graduate theses

Post-graduate theses

Search command : Author="Στεφανίδης" And Author="Κωνσταντίνος"

Current Record: 15 of 6511

[Add to Basket]

Identifier

000463769

Title

An automated method for the creation of oriented bounding boxes in remote sensing object detection datasets

Alternative Title

Μία μέθοδος για την αυτοματοποιημένη δημιουργία περιστραμμένων πλαισίων αντικειμένων σε εικόνες δορυφορικής τηλεπισκόπησης

Author

Σαβαθράκης, Γεώργιος Κων.

Thesis advisor

Αργυρός, Αντώνιος

Reviewer

Τραχανιάς, Παναγιώτης
Ζαμπούλης, Ξενοφών

Abstract

The detection of objects in remote sensing images is an important application of computer vision. Given the increasing demands regarding the surveillance and monitoring of areas that may include objects of interest, the task of accurately detecting objects from remote sensing aerial images is of signif- icant importance. Various object detection algorithms localize objects by identifying either their Horizontal Bounding Boxes (HBBs) or their Oriented Bounding Boxes (OBBs). OBBs provide a far more accurate/tighter local- ization of object regions as well as their orientation. Several object detection datasets provide annotations that include both HBBs and OBBs. However, many of them do not include OBB annotations. In this work, we propose a method which takes the objects’ HBB annotations as input, and automati- cally calculates the corresponding OBBs. The proposed method consists of three main parts, (a) object segmentation that is built upon the segment- anything model (SAM) to calculate object masks based on the information provided by the HBBs, (b) morphological filtering which eliminates possible artifacts stemming from the segmentation process, and (c) contour detection applied to the post-processed masks that are used to compute the optimal OBBs of the target objects. By automating the process of OBB annotation, the proposed method permits the exploitation of existing HBB-annotated datasets to train object detectors of improved performance. Furthermore, we propose the development of two data augmentation methods that resolve the problem of the objects’ orientation imbalance. We do this by either maintaining or increasing the number of objects in the dataset. Creating augmented datasets can lead to less biased datasets which when used for training can lead to more precise detections. We support this finding by re- porting the results of several experiments that involve three standard remote sensing object detection datasets, as well as state of the art oriented object detectors.

Language

English

Subject

Object detection

Oriented bounding boxes

Ανίχνευση αντικειμένων

Περιστραμμένα πλαίσια

Issue date

2024-03-22

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses

Type of Work--Post-graduate theses