This repository contains the code for the WACV 2025 paper Semantic Clustering of Image Retrieval Databases used for Visual Localization. The code can be used to adapt the procedure to other datasets and to reproduce our results after publishing the entire 4 Seasons dataset. The individual submodules that realize the individual components of our pipeline are collected here.
The data used within our experiments can be accessed using this link.
In our experiments, we use NetVlad to learn VPR descriptors which form a image retrieval database. For implementation, we used this repository, which we adapted to our dataset and managed in pytorch-NetVlad.
For semantic clustering we used a script for learning (semantic) clusters based on the semantic distributions within the semantic segmentation masks. Based on the learned clusters, we (soft-)assign query captures to search spaces and match VPR descriptors within the reduced search space. Our scripts for this are managed in sem-clustering.
Results of semantic segmentation are used as input for semantic clustering. For this, we trained a model that segments aerial captures in grayscale into basic forms of land cover. Our code for this is managed in sem-training.