Citation: Ruiz-Ponce, P.; Ortiz-Perez,
D.; Garcia-Rodriguez, J.; Kiefer, B.
POSEIDON: A Data Augmentation
Tool for Small Object Detection
Datasets in Maritime Environments.
Sensors 2023, 23, 3691. https://
doi.org/10.3390/s23073691
Academic Editor: Amir
Atapour-Abarghouei
Received: 19 February 2023
Revised: 30 March 2023
Accepted: 30 March 2023
Published: 2 April 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Article
POSEIDON: A Data Augmentation Tool for Small Object
Detection Datasets in Maritime Environments
Pablo Ruiz-Ponce
1,
* , David Ortiz-Perez
1
, Jose Garcia-Rodriguez
1
and Benjamin Kiefer
2
1
Department of Computer Technology and Computation, University of Alicante,
03690 San Vicente del Raspeig, Spain
2
Faculty of Science, University of Tuebingen, 72076 Tuebingen, Germany
* Correspondence: pruiz@dtic.ua.es
Abstract:
Certain fields present significant challenges when attempting to train complex Deep
Learning architectures, particularly when the available datasets are limited and imbalanced. Real-
time object detection in maritime environments using aerial images is a notable example. Although
SeaDronesSee is the most extensive and complete dataset for this task, it suffers from significant
class imbalance. To address this issue, we present POSEIDON, a data augmentation tool specifically
designed for object detection datasets. Our approach generates new training samples by combining
objects and samples from the original training set while utilizing the image metadata to make
informed decisions. We evaluate our method using YOLOv5 and YOLOv8 and demonstrate its
superiority over other balancing techniques, such as error weighting, by an overall improvement of
2.33% and 4.6%, respectively.
Keywords:
data augmentation; object detection; data imbalance; YOLO; aerial images; maritime
environments
1. Introduction
Although object detection has been extensively researched, with a plethora of trained
models and architectures available [
1
], there remain certain areas where large datasets
capable of training the most complex deep learning architectures are still lacking. One
of these areas pertains to the real-time detection of small vessels, individuals, and other
objects in maritime environments using aerial images obtained from drones or small aircraft.
Developing robust and precise models for this application would prove to be highly
beneficial in search and rescue missions, humanitarian aid efforts, and surveillance [
2
–
4
]
and security operations. However, as we have previously noted in our publication [
5
], the
primary issue is the high cost of capturing such images and the fact that instances in these
images tend to be very small. Additionally, maritime environments provide a large factor
of variance that has to be dealt with, and variations among the instances to be detected are
equally significant.
Our prior research focused on the SeaDronesSee dataset [
6
], which was introduced
in 2021 and encompasses a significant number of aerial images suitable for a variety of
applications, including object detection and object tracking. Although additional datasets,
such as Seagull [
7
], are also accessible, SeaDronesSee offers a greater diversity of images
with varying attributes, thereby permitting more thorough analysis. Nevertheless, the
primary challenge we encountered in our earlier work, and the one we address in this
paper, is the significant class imbalance among the various categories present in the dataset,
as exemplified in Figure 1.
The motivation of this research is to address the issue of dataset imbalance by examin-
ing a range of strategies and proposing a method for generating new samples to reduce
the imbalance. Due to the scarcity of available data, the focus has been on utilizing the
existing instances in the training set and leveraging image metadata to generate coherent
Sensors 2023, 23, 3691. https://doi.org/10.3390/s23073691 https://www.mdpi.com/journal/sensors