---
license: mit
language:
- en
base_model:
- facebook/dinov3-vits16-pretrain-lvd1689m
pipeline_tag: image-segmentation
tags:
- computer-vision
- image-segmentation
- vision-transformer
- marine-biology
- benthic-ecology
- pytorch
---

# SeaDino-Seg-1: Benthic Algae Segmentation Model

**SeaDino-Seg-1** is a specialized machine learning pipeline designed to automate benthic reef mapping and marine ecological survey analysis. It provides dense, pixel-level semantic segmentation of key ecological substrates and marine flora on rocky reefs, transforming raw underwater footage into structured ecological data.

*   **Repository Type:** Model Card / Weight Hub
*   **Target Domain:** Marine Benthic Ecology
*   **Base Architecture:** Vision Transformer (DINOv3 ViT-S/16)

---

## 📊 Model Variants

This repository hosts two distinct trained model configurations that can be evaluated independently or compared side-by-side:

1.  **SeaDino-Seg-1-Org (Model Org):** A standard baseline configuration optimized for rapid, general benthic feature extraction.
2.  **SeaDino-Seg-1-Fg (Model Fg):** A highly specialized configuration featuring a custom backbone optimized for low-contrast boundaries and high-density marine life identification.

Both models support multiple, interchangeable decoder head sizes:
*   `Tiny` (1D Linear Head)
*   `Small` (2D Spatial Conv Head)
*   `Medium` (3D Spatial Conv Head)
*   `Big` (4D Spatial Conv Head)

---

## 📁 Repository Files

This repository contains the following deployment-ready weights and configuration files:

*   `SeaDino-Seg-1-Fg-Backbone.ckpt` (The custom, specialized backbone weights)
*   `SeaDino-Seg-1-Fg-Small.pth` (The spatial decoder head weights for Model Fg)
*   `SeaDino-Seg-1-Org-Small.pth` (The spatial decoder head weights for Model Org)
*   `class_map.json` (The configuration file defining the benthic classes)

---

## 📈 Target Benthic Classes

The system is trained to identify and segment the following 6 ecologically vital classes:

| Class ID | Class Name | Color Code (RGB) | Description |
| :---: | :--- | :---: | :--- |
| **0** | Background | `[0, 0, 0]` | Barren sand, open water column, or unlabeled substrates |
| **1** | Rock | `[204, 51, 51]` | Exposed, barren rocky reef substrate |
| **2** | Carpophyllum | `[51, 204, 51]` | Canopy-forming brown algae (*Carpophyllum maschalocarpum*) |
| **3** | Ecklonia | `[204, 204, 51]` | Common kelp forest canopy (*Ecklonia radiata*) |
| **4** | Amphiroa | `[51, 51, 204]` | Articulated coralline algae (*Amphiroa anceps*) |
| **5** | Anthothoe | `[204, 51, 204]` | Encrusting white-striped anemone (*Anthothoe albocincta*) |

---

## ⚙️ How to Load and Use

To run predictions using these cloud weights, use our official [SeaDino-Seg-1 Github Pipeline](https://github.com/NeelJani1/Algae_webproject).

The pipeline automatically fetches these weights, manages memory safety on GPU/CPU, and outputs both visual maps (confidence heatmaps & comparative grids) and structured JSON percent-cover data.

### Example CLI Command:
```bash
python evaluate.py \
    --run_base \
    --run_ft \
    --sizes small \
    --mode all \
    --hf_repo "Neel536/Algea_Segmentation_Model"
```

---

## 📝 Citation & License

*   **License:** MIT
*   **Backbone Reference:** Self-Supervised ViT-S/16 (DINOv3)
*   **Project Lead:** Neel / SeaDino-seg-1