---
license: mit
task_categories:
- feature-extraction
language:
- en
tags:
- FAISS
- EarthLoc2
- Dinov2
size_categories:
- 100K<n<1M
---


# 🌍 EarthLoc2 FAISS Index (DINOv2 + SALAD)

This is the FAISS index used in the Hugging Face Space:  
🔗 [EarthLoc2: Image Geolocalization](https://huggingface.co/spaces/pawlo2013/EarthLoc2)

---

## 📦 What’s Inside?

This index was constructed by encoding all of the **2021 Sentinel-2 satellite images** from the EarthLoc dataset:
👉 [EarthLoc_2021_Database](https://huggingface.co/datasets/pawlo2013/EarthLoc_2021_Database)

Each image is encoded **four times** (with rotations of 0°, 90°, 180°, and 270°), and descriptors were extracted using:

https://huggingface.co/pawlo2013/EarthLoc2

Also see the corresponding faiss_index_webdataset.csv file which is used to map each index to the correct local path or shard path from [EarthLoc_2021_Database](https://huggingface.co/datasets/pawlo2013/EarthLoc_2021_Database). 

### 📊 Index Stats

| Metric                | Value       |
|-----------------------|-------------|
| Total descriptors     | ~600,000    |
| Descriptor dimension  | 3072        |
| FAISS index size      | ~8GB        |
| Image rotations       | 0°, 90°, 180°, 270° |

---

## 🧪 How to Use

You can download the trained model from https://huggingface.co/pawlo2013/EarthLoc2

### 🔧 1. Preprocess Query Image

```python
import torchvision.transforms as tfm

image_transform = tfm.Compose([
    tfm.Resize((model.image_size, model.image_size), antialias=True),
    tfm.ToTensor(),
    tfm.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

image_tensor = image_transform(query_image)

with torch.no_grad():
    descriptor = model(image_tensor.unsqueeze(0))  # Shape: (1, 3072)
descriptor_np = descriptor.cpu().numpy(

distances, indices = faiss_index.search(descriptor_np, num_results)
flat_indices = indices.flatten()


import torch.nn.functional as F

for faiss_idx in flat_indices:
    image_index = faiss_idx % num_db_images
    best_rotation_index = faiss_idx // num_db_images

    # Correct query image rotation (optional)
    query_tensor = F.rotate(query_tensor, [0, -90, -180, -270][best_rotation_index])