mever-team / visloc-estimation

Authors official PyTorch implementation of the "Leveraging EfficientNet and Contrastive Learning for Accurate Global-scale Location Estimation" [ICMR 2021] and "Leveraging Selective Prediction for Reliable Image Geolocation" [MMM 2022]

Visual location estimation

This repository contains the PyTorch implementation of the papers Leveraging EfficientNet and Contrastive Learning for Accurate Global-scale Location Estimation and Leveraging Selective Prediction for Reliable Image Geolocation. It provides all necessary code and pre-trained models for the evaluation of the location estimation method on the datasets presented in the paper to facilitate the reproducibility of the results. It provides code for the Search within Cell (SwC) scheme for the accurate location prediction, and the calculation of the Prediction Density (PD) that indicates image localizability.

Prerequisites

• Python 3
• PyTorch
• Torchvision

Preparation

Installation

• Clone this repo:

git clone https://github.com/mever-team/visloc-estimation
cd visloc-estimation

• Install all the dependencies by

pip install -r requirements.txt

Background collection

• Download and store the features of the background collection needed for SwC inference from here or by

wget https://mever.iti.gr/visloc/back_coll_features.hdf5

Evaluation datasets

• Follow the instructions in the original repos to download the evaluation datasets
  • Im2GPS
  • Im2GPS3k,YFCC4k
  • YFCC25k

Evaluation

• To evaluate the approach, run the evaluation.py script given the dataset you wand to evaluate on, the image_folder directory of the dataset images, and the background path of the background collection features

python evaluation.py  --dataset im2gps3k --image_folder <path_to_images> --background ./back_coll_features.hdf5

• The script above calculates the geolocation accuracy for the provided evaluation_radius. It provides results for the entire dataset and the dataset split of localizable and non-localizable based on the provided PD value.

• Change the values of top_k and eps to parametrize DBSCAN clustering of the SwC process.

• Change the values of conf_thres and conf_scale to parametrize the PD that determines which images are considered localizable and non-localizable.

Inference

• You can also use the code to predict the location of a single image

• Run the inference.py script given the image_path of an image stored locally or the image_url of an image from the web

python inference.py --image_url 'https://upload.wikimedia.org/wikipedia/commons/thumb/a/a8/Tour_Eiffel_Wikimedia_Commons.jpg/200px-Tour_Eiffel_Wikimedia_Commons.jpg'

Citation

If you use this code for your research, please consider citing our papers:

@inproceedings{kordopatis2021leveraging,
  title={Leveraging efficientnet and contrastive learning for accurate global-scale location estimation},
  author={Kordopatis-Zilos, Giorgos and Galopoulos, Panagiotis and Papadopoulos, Symeon and Kompatsiaris, Ioannis},
  booktitle={Proceedings of the International Conference on Multimedia Retrieval},
  year={2021}
}

@inproceedings{panagiotopoulos2022leveraging,
  title={Leveraging Selective Prediction for Reliable Image Geolocation},
  author={Panagiotopoulos, Apostolos and Kordopatis-Zilos, Giorgos and Papadopoulos, Symeon},
  booktitle={Proceedings of the International Conference on MultiMedia Modeling},
  year={2022},
}