CoRef-DETR: Decoupled Global Context and Selective Boundary-Category Refinement for Underwater Instance Segmentation

The gist The paper introduces CoRef-DETR, a framework for underwater instance segmentation that decouples global context and selective boundary-category refinement to improve performance across various underwater datasets.

How it works

CoRef-DETR is built upon a DETR-style architecture enhanced with two main components: Decoupled Global Context and Selective Boundary-Category Refinement. This framework aims to address the limitations of existing methods by improving the handling of complex underwater visual data through these specialized modules The paper proposes this approach to enhance segmentation quality in challenging underwater environments

The core mechanism involves decoupling the global context processing from the fine-grained boundary and category refinement steps This decoupling is intended to allow the model to better capture both broad scene understanding and precise instance delineation

Key Components and Techniques

The framework incorporates specific strategies to refine the segmentation output These refinements are designed to be selective rather than uniformly applied across all features

One of the key techniques is the Selective Boundary-Category Refinement module This module focuses on refining boundaries and categories specifically where they are most needed

Evaluation and Performance

The effectiveness of CoRef-DETR is demonstrated through extensive evaluation across multiple underwater instance segmentation datasets The paper reports results on various protocols including category-aware and class-agnostic salient instance tasks

The performance metrics reported include mAP, AP50, AP75, small object AP (APS), medium-object AP (APM), and large-object AP (APL) The results show competitive performance against prior state-of-the-art works across the tested categories

Conclusion

CoRef-DETR successfully integrates decoupled global context and selective boundary-category refinement into a DETR architecture This results in improved segmentation performance on underwater instance segmentation tasks across different scales and complexities The framework provides a robust solution for achieving high-quality instance segmentation in the underwater domain

The paper concludes that CoRef-DETR offers a strong method for enhancing underwater instance segmentation by selectively refining boundaries and categories within a decoupled global context setting. The framework is shown to be effective across various benchmarks, including those testing different generalization axes in marine perception. The proposed method provides a viable path toward achieving state-of-the-art results without relying on heavy foundation models or complex prompt generation during inference. The overall findings validate the efficiency and accuracy of this decoupled refinement strategy for underwater instance segmentation. The work contributes a compact yet powerful solution to the challenges faced in marine resource exploration and ecological monitoring. The results on salient instance datasets confirm the method's ability to handle foreground object relevance effectively. The successful implementation validates the potential of selective refinement in improving segmentation fidelity under underwater conditions. The framework is presented as a significant step forward in developing lightweight and effective underwater instance segmentation models. The paper demonstrates that targeted refinement can yield substantial gains in object-scale metrics across different image area protocols. The methodology is designed to be efficient while maintaining high accuracy in a domain where computational resources are often constrained. The study provides valuable insights into optimizing the trade-off between model complexity and segmentation quality in underwater vision tasks. The contribution lies in providing a practical framework that avoids the overhead associated with large foundation models while maintaining strong performance benchmarks. The proposed architecture is designed to be compact, making it suitable for deployment in real-time underwater perception systems. The overall study confirms the viability of this decoupled approach for achieving superior results in underwater instance segmentation tasks. The paper offers a novel contribution to the field by focusing on selective refinement within a DETR-style detector. The framework addresses the practical tension between high accuracy and deployment feasibility in marine applications. The evaluation across four distinct protocols provides a comprehensive assessment of the model's generalization capabilities. The study concludes that CoRef-DETR is a strong candidate for lightweight underwater salient instance segmentation detection. The proposed technique successfully addresses the need for compact yet high-performing models in underwater perception systems. The paper establishes a new benchmark for lightweight segmentation methods in the underwater domain. The results strongly suggest that selective boundary and category refinement is a critical component for achieving superior performance in this domain. The paper provides a valuable tool for researchers working on underwater image analysis and perception systems. The framework successfully balances the need for global context awareness with precise local refinement in a single inference pass. The study confirms that this approach is highly effective on salient instance datasets like USIS10K and USIS16K. The paper's findings establish a new baseline for lightweight detection transformers in the context of underwater segmentation. The research offers a practical and efficient solution to the challenges inherent in underwater instance segmentation.

1. AquaBSAM Feature Modulation: The system can extract critical underwater cues by converting normalized RGB input into a cue tensor and applying bounded residual modulation to intermediate feature maps, as shown in Equation (2). This allows the model to adapt visual features directly to underwater conditions without relying on external prompt generation or auxiliary modality estimation during inference.

2. Lightweight Architecture for Real-Time Deployment: The framework achieves a latency of 4.31–6.34 milliseconds (ms) latency on an NVIDIA T4 GPU using TensorRT half-precision (FP16). This makes the system capable of real-time inference under an accessible reproduction setting, enabling deployment on resource-constrained underwater robotic platforms where heavy foundation models are infeasible.

3. Robust Salient Instance Segmentation Across Protocols: LUSIS-DETR demonstrates competitive performance across various protocols, including category-aware and salient-instance protocols on datasets like USIS16K. The paper shows that the model improves small-, medium-, and largeobject AP over prior baselines under the benchmark’s normalized image-area protocol, ensuring high accuracy regardless of object size.

4. Training-Only Auxiliary Supervision: The training objective incorporates auxiliary losses such as boundary loss, structure loss, mask-quality supervision, and Lovasz loss. These terms are explicitly stated as not evaluated at inference, meaning the model learns to produce high-quality masks by supervising these specific metrics during training without incurring any computational overhead during deployment.