clip-camping-v2

CLIP v2 Camping Spot Detector

Fine-tuned CLIP model for identifying suitable camping locations from satellite imagery.

Model Details

• Base Model: OpenAI CLIP ViT-B/32
• Training Dataset: 3,161 satellite images (70% train, 15% val, 15% test)
• Training Duration: 30 epochs on Tesla T4 GPU
• Architecture: CLIP vision encoder + custom binary classification head
• Input: Satellite image (base64 encoded)
• Output: Campability score (0-100)

Performance Metrics

• Test Accuracy: 93.07%
• Precision: 95.35%
• Recall: 96.53%
• F1-Score: 95.94%

Performance Breakdown

• Campable spots detection: 96.53% recall
• False positive rate: 4.65%

Usage

Python (Local)

import torch
from PIL import Image
from transformers import CLIPProcessor, CLIPModel

# Load model and processor
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
checkpoint = torch.load("clip_v2_best.pth")

# Load image
image = Image.open("satellite_image.jpg")

# Process and classify
inputs = processor(images=image, return_tensors="pt")
with torch.no_grad():
    image_features = model.get_image_features(**inputs)
    classifier_output = classifier(image_features)
    score = torch.softmax(classifier_output, dim=1)[0][1]  # Campable probability

print(f"Campability Score: {score * 100:.1f}%")

Hugging Face Inference API

import requests
import base64

API_URL = "https://api-inference.huggingface.co/models/Soviet-Bear/clip-camping-v2"
headers = {"Authorization": f"Bearer {HF_TOKEN}"}

with open("image.jpg", "rb") as f:
    data = f.read()

response = requests.post(API_URL, headers=headers, data=data)
result = response.json()

Training Configuration

• Batch size: 32
• Learning rate: 0.001 (Adam optimizer)
• Epochs: 30
• Warmup: First 3 epochs
• Image size: 224x224
• Feature dimension: 512

Dataset

Training data sourced from:

• Park4Night camping locations (positive samples)
• OSM terrain unsuitable for camping (negative samples)
• Geographic distribution across multiple continents

Model Files

• clip_v2_best.pth: Trained classifier weights
• config.json: Model configuration
• model_index.json: Pipeline specification
• training_results.json: Complete training metrics
• training_history.png: Training/validation curves

Limitations

• Optimized for satellite imagery at 224x224 resolution
• Trained on specific terrain types; may not generalize to all regions
• Requires clear satellite imagery (cloud cover reduces accuracy)

Citation

If you use this model, please cite:

@model{clip-camping-v2,
  title={CLIP v2 Camping Spot Detector},
  author={Soviet-Bear},
  year={2025}
}

License

This model derivative is based on OpenAI's CLIP model. See LICENSE for details.