Title: MariData: A Single-Step Unpaired Image Translation Approach for Maritime Settings

Abstract: The advancement of robust perception systems for Maritime Autonomous Surface Ships (MASS) is significantly hindered by a lack of diverse training datasets, especially regarding adverse weather and low-light scenarios. Since acquiring paired images in the dynamic maritime environment is physically unfeasible, generating synthetic data through unpaired image-to-image translation presents a vital solution. Nevertheless, current generative models often fail to retain the intricate structural details of small navigational objects, a limitation caused by latent compression bottlenecks. This study presents a framework for producing synthetic maritime data utilizing CycleGAN-turbo, a one-step unpaired translation architecture. By integrating zero-convolution skip connections that circumvent the Variational Autoencoder (VAE) bottleneck, our method explicitly safeguards the details of small objects, such as distant vessels and sea marks, during the translation process. To train and evaluate models for Day-to-Foggy, Day-to-Sunset, and Day-to-Night domain translations, we assembled a dataset comprising 7,000 maritime images. Both qualitative assessments and variable-strength inference tests indicate that our technique successfully synthesizes realistic atmospheric conditions while preserving the scene’s underlying semantic structure. While the Day-to-Foggy and Day-to-Sunset models show strong structural retention, the Day-to-Night model reveals the issue of semantic hallucination—such as the creation of artificial coastal lights—stemming from unbalanced training distributions. Ultimately, this research establishes an efficient, structure-aware data synthesis pipeline that directly tackles the data scarcity bottleneck in autonomous maritime navigation.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC