---
license: apache-2.0
base_model:
 - SAM
 - ViT-B
 - PLAS
tags:
- image-segmentation
- foundation-model
- coral reef
- corals
- marine
datasets:
- CoralNet
---

<p align="center">
  <picture>
    <!-- <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/reefsupport/.github/main/profile/logo_text_horizontal_transparent_white.png"> -->
    <img src="https://raw.githubusercontent.com/reefsupport/.github/main/profile/logo_text_horizontal_transparent.png" alt="Reef Support Logo" width="40%">
  </picture>
</p>

# 🌊 Reef Support

We at **Reef Support** are dedicated to providing cutting-edge tools and technologies to support marine conservation efforts around the globe. Our mission is to protect and restore marine ecosystems using innovative AI-driven solutions. One key aspect is **coral segmentation**, which we integrate into our [**MariMap**](https://marimap.reef.support) application to empower marine biologists and conservationists with accurate and efficient coral reef analysis.

---

# CoralSCOP: Segment Any Coral Image on this Planet

[**GitHub**](https://github.com/zhengziqiang/CoralSCOP) | 
[**Project page**](https://coralscop.hkustvgd.com/) |
[**Paper**](https://coralscop.hkustvgd.com/CoralSCOP_files/CoralSCOP__Segment_any_COral_Image_on_this_Planet.pdf)

**CoralSCOP** is the first **foundation model** for coral reef segmentation. It builds upon [Segment Anything Model (SAM)](https://github.com/facebookresearch/segment-anything) but addresses the **semantic-agnostic** limitations and includes additional training data & strategies for underwater coral scenes. This powerful model will be made available through **Reef Support’s MariMap application**, enhancing our ability to analyze coral reefs worldwide.

## Model Weights
- **Checkpoint**: `vit_b_coralscop.pth` (ViT-B backbone)

## Main Features
- **Parallel semantic branch**: better ability to differentiate coral vs. non-coral, individual coral boundaries, coral species, etc.
- **Train with negatives**: reduce false positives in large-scale underwater imagery.
- **Promptable reef segmentation**: can accept point, box, or text prompts to refine or specify the target coral classes.

## Acknowledgement
Thank you to the whole [**CoralSCOP**](https://github.com/zhengziqiang/CoralSCOP) team for their amazing contribution to the Marine Science community 💙🌊