Official repository of "EP2P-Loc: End-to-End 3D Point to 2D Pixel Localization for Large-Scale Visual Localization".
We propose EP2P-Loc, a novel large-scale visual localization method that mitigates such appearance discrepancy and enables end-to-end training for pose estimation. This repository is built upon the foundations of Swin-Transformer, Fast Point Transformer, and EPro-PnP.
Updates
conda create -n ep2ploc python=3.6
conda activate ep2ploc
pip install -r requirements.txt conda install -c sirokujira python-pcl --channel conda-forge
## Dataset
### Download datasets
* [2D-3D-S and (Aligned) S3DIS](http://buildingparser.stanford.edu/dataset.html)
* [KITTI](https://www.cvlibs.net/datasets/kitti/eval_odometry.php)
### Preprocessing
cd datasets
python preprocess_2d3ds.py --data_path <2D-3D-S_path> --s3dis_path
python preprocess_kitti.py --data_path
## Training and Testing
TBU
## Citation
@INPROCEEDINGS{EP2PLoc2023ICCV, author = {Kim, Minjung and Koo, Junseo and Kim, Gunhee}, title = {EP2P-Loc: End-to-End 3D Point to 2D Pixel Localization for Large-Scale Visual Localization}, booktitle = {International Conference on Computer Vision (ICCV)}, year = {2023} }