HOSS ReID: Cross-Modal Ship Identification

• HOSS ReID is a cross-modal benchmark that matches vessel identities between optical and SAR images under varied imaging conditions.
• It enables rigorous evaluation of modality transfer using protocols like ALL→ALL, optical→SAR, and SAR→optical with metrics such as mAP and Rank-1 accuracy.
• The dataset supports innovative techniques such as diffusion-based synthesis and modality-consistent representation learning for enhanced retrieval performance.

The HOSS ReID dataset is a cross-modal ship re-identification (ReID) benchmark introduced to advance robust maritime target identification under severe optical-SAR (synthetic aperture radar) modality shifts. HOSS is leveraged as the primary empirical testbed in "MOS: Mitigating Optical-SAR Modality Gap for Cross-Modal Ship Re-Identification" (Zhao et al., 3 Dec 2025), where it enables rigorous evaluation of cross-domain ReID models under all-to-all, optical-to-SAR, and SAR-to-optical retrieval protocols. The dataset is critical for quantifying the substantial feature distribution discrepancy between optical and SAR domains and stimulating algorithmic advances in maritime surveillance, multi-sensor security, and domain adaptation research.

1. Dataset Definition and Scope

HOSS ReID comprises annotated optical and SAR imagery of maritime vessels, with the central task defined as matching ship identities across the two sensing modalities. The dataset is constructed to present fine-grained identity-level correspondence under large visual variations stemming from disparate imaging mechanisms and acquisition geometries. Its focus on real-world multi-modal maritime assets, documented under operational conditions, exposes models to spectrum-specific artifacts, noise, resolution variance, and occlusion challenges typical of naval and aerial surveillance.

2. Imaging Modalities and Modality Gap

The cornerstone of HOSS is its inclusion of both optical and SAR modalities per ship identity, enabling direct study of cross-modal representation learning and retrieval. The optical domain provides high-resolution, color-texture-rich views but suffers in poor visibility or at night. SAR images, in contrast, are robust to weather and illumination changes, encoding vessel structure via radar reflectivity, but exhibit severe speckle, geometric distortion, and a lack of fine visual cues.

This significant modality gap, quantified by highly non-overlapping feature distributions, motivates dedicated cross-domain alignment algorithms. HOSS supports evaluation protocols specifically isolating the Optical→SAR and SAR→Optical search directions.

3. Protocols and Evaluation Metrics

Three principal evaluation settings are defined:

• ALL → ALL: Query and gallery both consist of mixed optical and SAR images, measuring general retrieval robustness.
• Optical → SAR: Only optical images as queries and only SAR as gallery, directly testing modality-bridging capability.
• SAR → Optical: Inverse of above, highlighting difficulties in extracting discriminative features from SAR for matching against optical.

Performance metrics include mean Average Precision (mAP) and Rank-1 (R1) accuracy, adopted to facilitate direct comparison to established ReID literature and enable ablation-driven progress assessment.

4. Empirical Impact and Baseline Performance

HOSS is the reference benchmark in (Zhao et al., 3 Dec 2025) for evaluating state-of-the-art and proposed cross-modal ReID frameworks. The MOS pipeline demonstrates significant improvements on HOSS when compared to strong baselines:

• CDGF module alone delivers measurable mAP and R1 gains: ALL→ALL R1 +0.6%, Opt→SAR R1 +3.1%, SAR→Opt R1 +2.9%.
• With full MOS (CDGF+MCRL), the improvements are substantially larger, e.g., SAR→Opt R1 +16.4%.
• Full experimental results, including ablation of modules and feature fusion strategies, are tabulated to rigorously document algorithmic progress in all cross-modal scenarios (Zhao et al., 3 Dec 2025).

This scale and diversity of modality-resolved validation protocols are a distinguishing strength of HOSS relative to prior ReID datasets.

5. Algorithmic Innovations Enabled by HOSS

Because HOSS exposes the fundamental limitations of vanilla and even domain-adversarial ReID methods, it has catalyzed the following approaches:

• Diffusion-Based Cross-Modal Generation: The MOS method employs a Brownian-bridge diffusion model to synthesize pseudo-SAR images from optical queries, addressing feature mismatch at the sample level.
• Modality-Consistent Representation Learning (MCRL): The co-training of denoised SAR processing and class-wise alignment produces more modality-invariant embeddings.
• Convex Feature Fusion: The weighted ℓ₂-normalized average of original optical and K hallucinated SAR features delivers significant gains on HOSS, with optimal performance at τ=0.2 (Zhao et al., 3 Dec 2025).
• Comprehensive Ablation Analysis: The dataset enables fine-grained study of model components under stringent, real-world modality transfer demands.

6. Significance, Advances, and Limitations

HOSS defines the current benchmark for optical-SAR ship ReID, setting quantitative expectations for both single- and cross-modality scenarios in maritime surveillance. It supports detailed ablation and advancement of generative, alignment, and fusion methodologies critical for robust cross-domain retrieval in harsh operational environments.

A notable implication is that, by supporting template-based evaluation and multi-sample inference, HOSS provides an ideal substrate for future research on probabilistic retrieval, uncertainty quantification, and interpretable feature alignment in cross-modal maritime vision. However, expansion to include wider vessel classes, temporal sequences, and additional sensing modalities such as infrared or EO remains an open direction for the long-term generalization of cross-modal ReID methods.

---

For technical details and state-of-the-art results on HOSS ReID, see (Zhao et al., 3 Dec 2025).