Research map / Detection and segmentation / Open-set and few-shot detection
Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds USIS10K, a 10,632-image underwater salient instance segmentation dataset with 7 categories, and proposes USIS-SAM, a frozen-SAM-based model with adapter-based UA-ViT encoder and a Salient Feature Prompt Generator that outputs prompts without user input.
Problem
Underwater salient instance segmentation suffers from low accuracy because of complex underwater imaging conditions and the lack of large-scale pixel-level salient instance annotations; SAM needs manual point/box prompts and lacks underwater domain knowledge.
Approach
Adapt a frozen SAM ViT-H encoder with underwater-specific adapters and channel adapters (UA-ViT), and replace manual prompts with a Salient Feature Prompt Generator that fuses multi-layer, multi-scale features into prompt embeddings for end-to-end salient instance segmentation.
Key results
- USIS-SAM reaches 59.7 mAP class-agnostic and 43.1 mAP multi-class on USIS10K, outperforming OQTR (56.6) and WaterMask (58.3 class-agnostic) and RSPrompter (40.2 multi-class).
- On SIS10K, USIS-SAM reaches 70.1 mAP vs 67.2 for OQTR, but is weaker on AP75 (78.2 vs 81.7).
- On UIIS, USIS-SAM reaches 29.4 mAP vs 27.2 for WaterMask.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use USIS-SAM instead of the strongest listed baseline, WaterMask ResNet-101 | USIS10K class-agnostic segmentation · USIS10K class-agnostic | mAP: 59.0 -> 59.7 | Helped |
| Use USIS-SAM instead of RSPrompter | USIS10K multi-class segmentation · USIS10K multi-class | mAP: 40.2 -> 43.1 | Helped |
| Use USIS-SAM instead of OQTR | SIS10K segmentation · SIS10K | AP75: 81.7 -> 78.2 | Hurt |
| Use USIS-SAM instead of WaterMask | underwater instance segmentation · UIIS | mAP: 27.2 -> 29.4 | Helped |
| add UA-ViT (ablation) | USIS-SAM · USIS10K multi-class | mAP: 41.5 -> 43.1 | Helped |
| add SFPG instead of Multi-scale Feature Enhancer (ablation) | USIS-SAM · USIS10K multi-class | mAP: 42.2 -> 43.1 | Helped |
| add Adapter (ablation) | UA-ViT in USIS-SAM · USIS10K multi-class | mAP: 41.7 -> 43.1 | Helped |
| add Channel Adapter (ablation) | UA-ViT in USIS-SAM · USIS10K multi-class | mAP: 42.0 -> 43.1 | Helped |
| add Salient Feature Fusion Module (ablation) | SFPG in USIS-SAM · USIS10K multi-class | mAP: 42.3 -> 43.1 | Helped |
| add Multi-Up module (ablation) | SFPG in USIS-SAM · USIS10K multi-class | mAP: 42.5 -> 43.1 | Helped |
Surprising findings
- Only ~9.3% of USIS-SAM parameters are trainable, with SAM weights frozen.
- The authors attribute weaker AP75 on SIS10K to SAM's pretraining overlapping with the UA-ViT's learned knowledge, suggesting limited gain on land data.
Limitations
- USIS10K has relatively small local boundary contrast, making boundary segmentation harder.
- On SIS10K the method is weaker on AP75 than OQTR.
- Trained and evaluated on a single underwater dataset with 7 categories; generalization to other underwater domains is limited.
Experimental setup
Datasets: USIS10K SIS10K UIIS
Benchmarks: USIS10K SIS10K UIIS
Metrics: mAP AP50 AP75
Baselines: S4Net RDPNet OQTR WaterMask SAM+BBox SAM+Mask RSPrompter URank+RDPNet URank+OQTR URank+RSPrompter
Builds on: Segment Anything Model RSPrompter Mask R-CNN Parameter-efficient fine-tuning adapters
Closest papers by meaning
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.89)
- Fantastic Animals and Where to Find Them: Segment Any Marine Animal with Dual SAM (CVPR 2024 · similarity 0.87)
- NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding (NeurIPS 2025 · similarity 0.86)
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.86)
- WebUOT-1M: Advancing Deep Underwater Object Tracking with A Million-Scale Benchmark (NeurIPS 2024 · similarity 0.85)
- A Unified Image-Dense Annotation Generation Model for Underwater Scenes (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.