Research map / Detection and segmentation / CNN features
Fantastic Animals and Where to Find Them: Segment Any Marine Animal with Dual SAM
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.
Proposes Dual-SAM, a dual-branch SAM-based framework with multi-level coupled prompts, dilated fusion attention, criss-cross connectivity prediction and pseudo-label mutual supervision for marine animal segmentation.
Problem
Marine animal segmentation under underwater conditions, where CNN/Transformer methods lack long-range context and SAM, trained on natural images with single-position prompts, performs poorly.
Approach
Adapts SAM's encoder with LoRA and adapters in a dual (original + gamma-corrected) branch, injects multi-level auto-prompts coupled across both branches, fuses multi-level features with dilated fusion attention, and predicts 8-channel criss-cross connectivity maps supervised by mutual pseudo-labels between two decoders.
Key results
- On MAS3K, Dual-SAM reaches 0.789 mIoU and 0.023 MAE, versus 0.748 mIoU for H2Former and 0.742 for SAM-DA.
- On UFO120, Dual-SAM achieves 0.810 mIoU, above the next best H2Former at 0.780.
- On USOD10K, Dual-SAM achieves S-alpha 0.9238 and maxF 0.9311, outperforming TC-USOD (0.9215 S-alpha).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace pixel-wise prediction with C3P (ablation) | Dual-SAM · MAS3K | mIoU: 0.772 -> 0.789 | Helped |
| replace Nearby connectivity prediction with C3P (ablation) | Dual-SAM · MAS3K | mIoU: 0.781 -> 0.789 | Helped |
| use dual branches instead of a single branch (ablation) | Dual-SAM · MAS3K | mIoU: 0.767 -> 0.789 | Helped |
| add pseudo-label mutual supervision (ablation) | Dual-SAM dual branches · MAS3K | mIoU: 0.771 -> 0.789 | Helped |
| add Multi-level Coupled Prompt (ablation) | Dual-SAM · MAS3K | mIoU: 0.778 -> 0.789 | Helped |
| add Dilated Fusion Attention Module (ablation) | Dual-SAM · MAS3K | mIoU: 0.769 -> 0.789 | Helped |
| add LoRA (ablation) | Dual-SAM · MAS3K | mIoU: 0.768 -> 0.789 | Helped |
| add Adapter (ablation) | Dual-SAM · MAS3K | mIoU: 0.774 -> 0.789 | Helped |
Surprising findings
- Plain SAM performs poorly on marine data (0.566 mIoU on MAS3K), well below CNN and Transformer baselines.
Limitations
- Plain SAM is far weaker than task-specific methods on marine data, so the framework depends on SAM-B adaptation.
- Gains over strongest baselines are modest on some datasets (e.g., RUWI mIoU 0.904 vs 0.871).
Experimental setup
Datasets: MAS3K RMAS UFO120 RUWI USOD10K
Benchmarks: MAS3K RMAS UFO120 RUWI USOD10K
Metrics: mIoU S-alpha weighted F-measure mean E-phi MAE maxF
Baselines: SINet PFNet RankNet C2FNet ECDNet OCENet ZoomNet MASNet SETR TransUNet H2Former SAM
Builds on: Segment Anything Model LoRA Nearby connectivity prediction SAM-Adapter SAM-DA
Closest papers by meaning
- Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset (ICML 2024 · similarity 0.87)
- Fish-Vista: A Multi-Purpose Dataset for Understanding & Identification of Traits from Images (CVPR 2025 · similarity 0.84)
- SeafloorAI: A Large-scale Vision-Language Dataset for Seafloor Geological Survey (NeurIPS 2024 · similarity 0.84)
- LAqua: Laplacian Pyramids for Aquatic Segmentation (Student Abstract) (AAAI 2025 · similarity 0.83)
- CoralSCOP: Segment any COral Image on this Planet (CVPR 2024 · similarity 0.83)
- Enhancing Vision-Language Models with Morphological and Taxonomic Knowledge: Towards Coral Recognition for Ocean Health (AAAI 2025 · similarity 0.82)
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.