Boundary-Aware Refinement with Environment-Robust Adapter Tuning for Underwater Instance Segmentation

Pin-Chi Pan, Soo-Chang Pei

Abstract: Underwater instance segmentation is a challenging task due to adverse visual conditions such as light attenuation, scattering, and color distortion, which severely degrade image quality and hinder model performance. In this work, we propose \textbf{BARD-ERA}, a unified framework that integrates three novel components to address these challenges. First, the \textbf{Boundary-Aware Refinement Decoder (BARDecoder)} improves mask quality through progressive feature refinement and lightweight upsampling using a Multi-Stage Gated Refinement Network and Depthwise Separable Upsampling. Second, the \textbf{Environment-Robust Adapter (ERA)} enables efficient adaptation to underwater degradations by injecting environment-specific priors with over 90\% fewer trainable parameters than full fine-tuning. Third, the \textbf{Boundary-Aware Cross-Entropy (BACE) loss} enhances boundary supervision by leveraging range-null space decomposition. Together, these modules achieve state-of-the-art performance on the UIIS dataset, surpassing Mask R-CNN by 3.4 mAP with Swin-B and 3.8 mAP with ConvNeXt V2-B, while maintaining a compact model size. Our results demonstrate that BARD-ERA enables robust, accurate, and efficient segmentation in complex underwater scenes.

Submission Number: 151