EP2P-Loc: End-to-End 3D Point to 2D Pixel Localization for Large-Scale Visual Localization (ICCV 2023)

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

• Aug 18, 2023: Release benchmark datasets
• Jul 18, 2023: Initial commit

Requirements

• Ubuntu 16.04
• Python 3.6

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
• KITTI

Preprocessing

cd datasets

# 2D-3D-S
python preprocess_2d3ds.py --data_path <2D-3D-S_path> --s3dis_path <S3DIS_path> --cache_path <cache_path(optional)> --save_path <save_path>

# KITTI
python preprocess_kitti.py --data_path <KITTI_path> --save_path <save_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}
}