File size: 3,367 Bytes
2e5dabb 4d551c4 2e5dabb 4d551c4 2a6f55d 4d551c4 9be19fb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 | ---
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 |