University of Twente Student Theses

Login

Impact of Phase Data Augmentation on Performance and Robustness in Object Detection

de Witte, S.W. (2025) Impact of Phase Data Augmentation on Performance and Robustness in Object Detection.

[img] PDF
1MB
Abstract:When an image is converted to the Fourier domain, it can be represented by two components: amplitude and phase. Inherently, phase data captures object shapes, which are crucial for object detection. Therefore, applying data augmentation in the phase domain has the potential to enhance object detection models in ways unmatched by traditional augmentation methods. However, despite significant successes in image classification, the application of phase data augmentation in object detection is sparse. This research investigates the effectiveness of phase data augmentation methods in enhancing object detection performance and robustness. Additionally, a novel method applying phase data augmentation selectively within bounding boxes (BBoxes) is introduced, which reduces computational overhead by only applying the costly FFT operation on an area within bounding boxes. Results show that integrating phase data augmentation with default pipelines achieves slightly higher mean average precision (mAP) in general and improves robustness against noise corruptions, though limitations were observed under blur and other miscellaneous corruptions. It was also observed that phase data augmentation had the greatest effect on performance when applied to simpler models on smaller datasets. Future research directions include optimizing hyperparameters, exploring complementary augmentation strategies, and further refining BBox-specific augmentations for enhanced performance and robustness.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:https://purl.utwente.nl/essays/104933
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page