Object Detection in Maritime Surveillance: A Data-Centric Comparison of YOLO and Transformer-Based Models

This research focuses on improving maritime ob-ject detection by refining deep neural network models with the Roboflow ship detection dataset, which contains labeled images of ships. The main goal is to assess the performance of different YOLO architecture versions (YOLOv7 and YOLOv8) in comparison to the Detection Transformer (DETR), a model introduced by Facebook AI Research in 2020. Unlike traditional object detection approaches, which often utilize convolutional neural networks (CNNs) along with post-processing techniques such as anchor boxes or region proposal networks (RPNs), DETR simplifies the process by removing these additional components. In this study, both YOLO models and DETR were fine-tuned on the Roboflow dataset, and their performance was measured in detecting maritime objects. The findings indicate that while the YOLO models achieved strong results, DETR demonstrated competitive accuracy with a more streamlined process. This study underscores the promise of transformer-based models for specialized object recognition tasks, offering valuable insights for enhancing maritime detection systems.