RANGE: Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings (CVPR 2025) 🌎🌍🌏

This repository is the official implementation of RANGE. RANGE (Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings) is a retrieval-augmented framework for embedding geographic coordinates. RANGE directly estimates the visual features for a given location, allowing the representations to capture high-resolution information.

🔥 Multi-scale Geoembeddings

Our method enforces a spatial smoothness constraint. Manipulating this constraint allows generating geo-embeddings at desired frequencies.

🏋️‍♂️ Performance on Downstream Tasks

We showed through a large number of downstream tasks that RANGE embeddings outperform several state-of-the-art location embedding methods such as SatCLIP, GeoCLIP, CSP, etc.

⚙️ Usage

The required model weights and embeddings are made available in huggingface. You can download the precomputed RANGE database using huggingface-cli. Currently, there are two possible choices: range_db_large.npz and range_db_med.npz.

git clone git@github.com:mvrl/RANGE.git
cd RANGE
#this is optional as we can directly do this inside our python script
huggingface-cli download mvrl/RANGE-database range_db_large.npz \
  --repo-type dataset \
  --local-dir ./pretrained/range \
  --local-dir-use-symlinks False

💻 Compute RANGE embeddings using load_model

# Create a new python file: touch ./range/test.py
import os
import torch
from huggingface_hub import hf_hub_download
from sklearn.preprocessing import MinMaxScaler

#import load_model locally
from .load_model import load_model

device = 'cuda' if torch.cuda.is_available() else 'cpu'
#get path to the pretrained SatCLIP model
pretrained_path =  hf_hub_download('microsoft/SatCLIP-ViT16-L40', 'satclip-vit16-l40.ckpt',
                                        repo_type='model', local_dir='./pretrained/range', local_dir_use_symlinks=False)

#get path to the RANGE database
db_path = hf_hub_download('mvrl/RANGE-database', 'range_db_large.npz',
                             repo_type='dataset', local_dir='./pretrained/range', local_dir_use_symlinks=False)

#define the model you want to load
model_name = 'RANGE+'
#set the beta parameter
beta = 0.5

#initialize model using load_model
rangep_model = load_model(model_name=model_name, pretrained_path=pretrained_path,
                               device=device, db_path=db_path, beta=beta)

#create a random Nx2 tensor    
a = torch.rand(10000,2).double().to(device)

#generate embeddings
# For optimal performance, use a veryyy large batch size.
# We consistently used a batch size of 10000 or higher when computing embeddings.
locs = torch.rand(10000, 2).double().to(device)
embeddings = rangep_model(locs)
scaler = MinMaxScaler()
scaled_embeddings = scaler.fit_transform(embeddings)
print(scaled_embeddings.shape)

python -m range.test
------------------------------
Output: (10000, 1280)

The load_model module can be used to load other soTA location encoders such as SatCLIP, GeoCLIP, CSP, SINR, etc. Look inside ./range/load_model.py file for details on usage and which location encoders are currently supported.

📑 Citation

@article{dhakal2025range,
  title={RANGE: Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings},
  author={Dhakal, Aayush and Sastry, Srikumar and Khanal, Subash and Ahmad, Adeel and Xing, Eric and Jacobs, Nathan},
  booktitle={Computer Vision and Pattern Recognition},
  year={2025},
  organization={IEEE/CVF}
}

🔍 Additional Links

Check out our lab website for other interesting works on geospatial understanding and mapping:

• Multi-Modal Vision Research Lab (MVRL)
• Related Works from MVRL