From Open Waters to Enclosed Cabins: ProteusVPR for Cross-Scene Visual Place Recognition in Maritime Perception and Cabin Inspection

Researchers have made significant advances in Visual Place Recognition (VPR) and geo-localization, addressing challenges in maritime environments and urban settings.

Autonomous robotic inspection in maritime environments faces unique VPR challenges due to cross-scene perceptual shifts. Existing VPR methods fail to generalize across different scenarios^[1]. To address this, a two-stage retrieval-refinement framework called ProteusVPR has been proposed. It employs a standard VPR model for initial image retrieval and a geometric-visual estimation network for precise localization. The XHZ dataset, an 8K-panoramic ship-borne dataset, was introduced to support this task. ProteusVPR consistently improves localization accuracy across multiple VPR backbones, reducing mean localization error by over 60% on average^[1]. Meanwhile, recent advances in vision-language models (VLMs) have shown promise for geo-localization. Traditional retrieval-based methods struggle with scalability and perceptual aliasing, while classification-based approaches lack generalization. A proposed framework leveraging a VLM to guide retrieval search space has outperformed prior state-of-the-art methods at street and city levels^[2].