This repository contains a re-implementation of our paper Interpretable Semantic Photo Geolocation.

Subpackage semantic_partitioning contains:

• script for reverse geocoding
• raw dataset visualization
• scripts to construct the semantic partitioning (SemP)

See semantic_partitioning/README.md for details.

Subpackage geo_classification contains:

• script to train from scratch
• evaluation pipeline including testsets
• pretrained models (EfficientNet-B4):

See geo_classification/README.md for details.

To overcome the need for a full installation of a reverse geocoder such as Nominatim, we provide the postprocessed output of the reverse geocoding for the MP-16 dataset¹ along with the validation set (YFCC-Val26k) which originally comprising photos and respective GPS coordinates. Both datasets are subsets of the YFCC100M dataset² which are crawled from Flickr.

Further details: semantic_partitioning/README.md

We provide the underlying functionality to compute the presented concept influence metric based on given semantic maps and attribution/explanation maps. Please note, that the computation of both semantic maps and explanation maps are not part of this repository.

conda env create -f environment.yml
conda activate github_semantic_geo_partitioning
# cd in respective subpackages
cd semantic_partitioning
cd geo_classification

@InProceedings{Theiner_2022_WACV,
    author    = {Theiner, Jonas and M\"uller-Budack, Eric and Ewerth, Ralph},
    title     = {Interpretable Semantic Photo Geolocation},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2022},
    pages     = {750-760}
}

This work is published under the GNU GENERAL PUBLIC LICENSE Version 3, 29 June 2007. For details please check the LICENSE file in the repository.