|
|
--- |
|
|
language: "en" |
|
|
license: "gpl-3.0" |
|
|
tags: |
|
|
- segmentation |
|
|
- computer-vision |
|
|
- yolo |
|
|
- beach |
|
|
- water |
|
|
- open-source |
|
|
task_categories: |
|
|
- image-segmentation |
|
|
--- |
|
|
|
|
|
# 🌊 Water Surface Segmentation on Beach Images |
|
|
|
|
|
## Model Overview |
|
|
This model performs **semantic segmentation of water surfaces** in beach or coastal images. |
|
|
It’s a fine-tuned version of **YOLOv11n**, adapted for **binary segmentation** with a single class: **`water`**. |
|
|
|
|
|
Built for lightweight, real-time deployment, the model achieves strong accuracy while remaining small and efficient. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧠 Model Details |
|
|
- **Architecture**: YOLOv11n segmentation head (binary) |
|
|
- **Base framework**: PyTorch / Ultralytics YOLOv11 |
|
|
- **Input size**: 640×640 RGB images |
|
|
- **Output**: Binary segmentation mask (1 class — `water`) |
|
|
- **Model file**: `nwsd-v2.pt` (≈6 MB) |
|
|
|
|
|
--- |
|
|
|
|
|
## 🚀 Key Features |
|
|
- ⚡ **Real-time inference** on CPU/GPU |
|
|
- 🖼 **Outputs**: Binary masks, overlays, and coverage statistics |
|
|
- 📊 **Evaluation tools** included for metrics & visualization |
|
|
- 🐍 **Easy Python integration** via a simple API (`nwsd_api.py`) |
|
|
|
|
|
--- |
|
|
|
|
|
## 📈 Performance |
|
|
| Metric | Value | Notes | |
|
|
|:--|:--|:--| |
|
|
| **mAP50** | > 0.85 | On validation set | |
|
|
| **Inference speed** | ~50 ms/image | On CPU | |
|
|
| **GPU memory** | < 2 GB | For 640×640 input | |
|
|
|
|
|
--- |
|
|
|
|
|
## 🗂 Dataset |
|
|
- **Type**: Binary segmentation |
|
|
- **Classes**: `water` |
|
|
- **Annotations**: PNG masks |
|
|
- **Source**: Custom-labeled beach dataset |
|
|
|
|
|
🔗 [Dataset on Roboflow](https://universe.roboflow.com/neptune-uxxqf/neptune-water-surface-detection) |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧩 Intended Uses |
|
|
**Use cases:** |
|
|
- Coastal or maritime monitoring |
|
|
- Beach safety & drowning prevention systems |
|
|
- Environmental analysis (e.g., water coverage estimation) |
|
|
|
|
|
**Limitations:** |
|
|
- Designed for daylight, clear beach imagery |
|
|
- May underperform in low-visibility or night-time scenes |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧪 How to Use |
|
|
|
|
|
### Load model from Hub |
|
|
```python |
|
|
from huggingface_hub import hf_hub_download |
|
|
import torch |
|
|
|
|
|
model_path = hf_hub_download(repo_id="Ehlum-Lucas/NWSD", filename="nwsd-v2.pt") |
|
|
model = torch.load(model_path, map_location="cpu") |
|
|
model.eval() |
|
|
``` |
|
|
|
|
|
### Inference example |
|
|
```python |
|
|
from PIL import Image |
|
|
import torch |
|
|
from torchvision import transforms |
|
|
|
|
|
img = Image.open("beachTest.jpg").convert("RGB") |
|
|
input_tensor = transforms.ToTensor()(img).unsqueeze(0) |
|
|
|
|
|
with torch.no_grad(): |
|
|
pred = model(input_tensor) |
|
|
``` |
|
|
|
|
|
### ⚙️ Training |
|
|
You can fine-tune or retrain the model using YOLOv11 tools: |
|
|
```bash |
|
|
python train.py --data data.yaml --weights <path_to_weights> --img 640 --batch 16 --epochs 50 |
|
|
``` |
|
|
Example configuration (data.yaml) defines paths to your datasets and class names. |
|
|
|
|
|
### 🧭 Evaluation |
|
|
```bash |
|
|
python evaluate.py --data data.yaml --weights model/nwsd-v2.pt |
|
|
``` |
|
|
|
|
|
Generates: |
|
|
|
|
|
- Binary mask |
|
|
- Overlay visualization |
|
|
- Water coverage stats |
|
|
|
|
|
## License |
|
|
This model is released under the **GPL-3.0 License**. See the [LICENSE](LICENSE) file for details. |
|
|
|
|
|
## Citation |
|
|
If you use this model in your work, please consider citing: |
|
|
```latex |
|
|
@misc{nwsd2025, |
|
|
title={Water Surface Segmentation on Beach Images}, |
|
|
author={Lucas Iglesia}, |
|
|
year={2025}, |
|
|
howpublished={\url{https://huggingface.co/Ehlum-Lucas/NWSD}} |
|
|
} |
|
|
``` |
|
|
|
|
|
|
