Research map / Detection and segmentation / Open-set and few-shot detection
CoralSCOP: Segment any COral Image on this Planet
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 CoralSCOP, a SAM-based coral segmentation foundation model with a parallel semantic branch, pre-trained on the new 41K-image CoralMask dataset for automatic, prompted, and few-shot semantic coral segmentation.
Problem
Dense coral reef segmentation requires heavy manual annotation by coral biologists; SAM is semantic-agnostic and over-segments corals, and existing coral datasets are small and narrow.
Approach
Extend SAM with a parallel semantic branch in the mask decoder that is trained with "non-coral" negative masks, enabling semantic coral masks, user-defined few-shot tuning, mask-referring and text-instruction segmentation, and sparse-to-dense conversion.
Key results
- With 1 point + bbox prompts, CoralSCOP (ViT-L) reaches 79.13 IoU on coral mask generation vs 72.80 for SAM ViT-H; automatic mode reaches 59.02 IoU with ViT-B vs 31.16 for SAM.
- Growth-form mIoU with 10-shot tuning rises from 32.54 (SegFormer) to 52.16 for CoralSCOP, and full-data 67.82 vs 59.22.
- On USOD10K salient segmentation, CoralSCOP (ViT-L) improves S-measure from 0.8843 to 0.8884 over SAM ViT-L.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add negative-mask training to CoralSCOP (ablation) | CoralSCOP ViT-L · Coral mask generation, automatic | IoU: 38.15 -> 46.46 | Helped |
| Add negative-mask training to CoralSCOP (ablation) | CoralSCOP ViT-L · Coral mask generation, automatic | MAE: 0.2601 -> 0.1814 | Helped |
| Add negative-mask training to CoralSCOP (ablation) | CoralSCOP ViT-L · Coral mask generation, 1 point + BBOXmask | IoU: 67.84 -> 68.66 | Helped |
| Add negative-mask training to CoralSCOP (ablation) | CoralSCOP ViT-L · Coral mask generation, 1 point + BBOXmask | MAE: 0.1475 -> 0.1408 | Helped |
| Use CoralSCOP instead of SegFormer for few-shot growth form segmentation | CoralSCOP · Growth form segmentation, 10-shot | mIoU: +19.62 | Helped |
| Use CoralSCOP instead of SegFormer for few-shot genus segmentation | CoralSCOP · Genus segmentation, 10-shot | mIoU: +5.58 | Helped |
| Use CoralSCOP for underwater salient object segmentation instead of SAM | CoralSCOP ViT-B · USOD10K | Sm: +0.0044 | Helped |
Surprising findings
- Fine-tuning SAM directly on coral masks yields many false positives, so the semantic branch with negatives is needed rather than plain fine-tuning.
- GroundingDINO+SAM for text-instructed coral segmentation cannot generate satisfactory masks.
Limitations
- Automatic mode mask quality is still well below prompted settings (e.g., 26.45 vs 66.78 IoU for ViT-B).
- Ablations trained for only one epoch due to compute constraints.
- Growth form and genus tuning use small datasets (500 and 400 images).
Experimental setup
Datasets: CoralMask USOD10K Corals of the World
Benchmarks: CoralMask (941-image test set) USOD10K growth form (100 test images) genus (100 test images) sparse-to-dense benthic test set (100 images)
Metrics: IoU pixel accuracy MAE mIoU prediction error Sm Emax maxF
Baselines: SAM SAM-Adapter Semantic-SAM SegFormer Mask2Former MaskFormer DeepLabV3 CPCe PLAS GroundingDINO+SAM
Builds on: SAM CLIP Semantic-SAM SAM-Adapter
Closest papers by meaning
- Enhancing Vision-Language Models with Morphological and Taxonomic Knowledge: Towards Coral Recognition for Ocean Health (AAAI 2025 · similarity 0.91)
- ReefNet: A Large-Scale Dataset and Benchmark for Fine-Grained Coral Reef Recognition (NeurIPS 2026 · similarity 0.90)
- CoralVQA: A Large-Scale Visual Question Answering Dataset for Coral Reef Image Understanding (NeurIPS 2025 · similarity 0.88)
- SeafloorAI: A Large-scale Vision-Language Dataset for Seafloor Geological Survey (NeurIPS 2024 · similarity 0.85)
- LAqua: Laplacian Pyramids for Aquatic Segmentation (Student Abstract) (AAAI 2025 · similarity 0.83)
- Fantastic Animals and Where to Find Them: Segment Any Marine Animal with Dual SAM (CVPR 2024 · similarity 0.83)
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.