Domain Adaptation from Generated Multi-Weather Images for Unsupervised Maritime Object Classification

TL;DR

The paper addresses unsupervised maritime object classification under long-tail distributions of object categories and weather conditions by introducing AIMO, a large AI-generated, weather-diverse dataset, and RMO, a real-world long-tail benchmark. It proposes a domain-adaptive framework that leverages AIMO as the source and unlabeled RMO as the target, combining CLIP-based generalization through Self-Knowledge Distillation, adversarial domain alignment, and classifier refinement with hidden-space perturbations, guided by a curriculum-based learning strategy. The approach yields significant improvements, achieving an average accuracy of approximately 73.7% on RMO and particularly boosting performance for rare categories and adverse weather scenarios, while providing ablations and parameter analyses to validate each component. The work advances practical maritime perception by enabling robust recognition across varied weather with limited labeled data, and it releases both the datasets and code for community use.

Abstract

The classification and recognition of maritime objects are crucial for enhancing maritime safety, monitoring, and intelligent sea environment prediction. However, existing unsupervised methods for maritime object classification often struggle with the long-tail data distributions in both object categories and weather conditions. In this paper, we construct a dataset named AIMO produced by large-scale generative models with diverse weather conditions and balanced object categories, and collect a dataset named RMO with real-world images where long-tail issue exists. We propose a novel domain adaptation approach that leverages AIMO (source domain) to address the problem of limited labeled data, unbalanced distribution and domain shift in RMO (target domain), enhance the generalization of source features with the Vision-Language Models such as CLIP, and propose a difficulty score for curriculum learning to optimize training process. Experimental results shows that the proposed method significantly improves the classification accuracy, particularly for samples within rare object categories and weather conditions. Datasets and codes will be publicly available at https://github.com/honoria0204/AIMO.

Figures (4)

• Figure 1: Examples and data statistics of AIMO and RMO. (a): Examples of AIMO with multiple weather and illumination conditions. (b): Examples of RMO with multiple weather and illumination conditions. (c): The number of AIMO and RMO with multiple weather and illumination conditions. (d): The number of AIMO with different categories. (e): The number of RMO with different categories.
• Figure 2: An overview of the proposed method. We use a series of Vision Transformer Blocks as backbone of feature extraction and take labels (classes and weather conditions) and images from the source domain, as well as images from the target domain as inputs. The designed UDA framework for maritime object classification is an adversarial adaptation network, consisting of generalization enhancement for source features, adversarial adaptation from source to target features and classifier refinement with perturbed features from both domains.
• Figure 3: Validation experiments with our proposed method and SCAN. (a): Comparison of accuracy on RMO with multiple weather & illumination conditions. (b): Comparison of accuracy on RMO with different categories.
• Figure 4: Parameter sensitivity experiments.