POSEIDON: A Data Augmentation Tool for Small Object Detection Datasets in Maritime Environments.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Training deep learning models for object detection is challenging with limited and imbalanced datasets.
• Real-time maritime object detection using aerial imagery, like the SeaDronesSee dataset, faces significant class imbalance issues.

Purpose of the Study:

• To develop a data augmentation tool, POSEIDON, to mitigate class imbalance in object detection datasets.
• To improve the performance of deep learning models in challenging environments like maritime settings.

Main Methods:

• POSEIDON generates new training samples by combining objects and samples from existing datasets.
• It leverages image metadata for informed data augmentation decisions.
• The method was evaluated using YOLOv5 and YOLOv8 object detection architectures.

Main Results:

• POSEIDON demonstrated superior performance compared to other balancing techniques.
• An overall improvement of 2.33% with YOLOv5 and 4.6% with YOLOv8 was achieved.

Conclusions:

• POSEIDON effectively addresses class imbalance in object detection datasets.
• The tool enhances the performance of deep learning models for maritime object detection.