Point-Line to Map Regresssion for Camera Relocalization

Project Page | PL2Map | PL2Map++

Introduction

We introduce a lightweight neural network for visual localization that efficiently represents both 3D points and lines. Specifically, we use a single transformer block to convert line features into distinctive point-like descriptors. These features are then refined through self- and cross-attention in a graph-based framework before 3D map regression using simple MLPs. Our method outperforms Hloc and Limap in small-scale indoor localization and achieves the best results in outdoor settings, setting a new benchmark for learning-based approaches. It also operates in real-time at ~16 FPS, compared to Limap’s ~0.03 FPS, while requiring only lightweight network weights of 33MB instead of Limap’s multi-GB memory footprint.

---

Papers

Improved 3D Point-Line Mapping Regression for Camera Relocalization
Bach-Thuan Bui, Huy-Hoang Bui, Yasuyuki Fujii, Dinh-Tuan Tran, and Joo-Ho Lee.
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025.
pdf

Representing 3D sparse map points and lines for camera relocalization
Bach-Thuan Bui, Huy-Hoang Bui, Dinh-Tuan Tran, and Joo-Ho Lee.
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024.
pdf

Installation

Python 3.9 + required packages

git clone https://github.com/ais-lab/pl2map.git
cd pl2map
git submodule update --init --recursive
conda create --name pl2map python=3.9
conda activate pl2map
# Refer to https://pytorch.org/get-started/previous-versions/ to install pytorch compatible with your CUDA
python -m pip install torch==1.12.0 torchvision==0.13.0 
python -m pip install -r requirements.txt

Supported datasets

• Microsoft 7scenes
• Cambridge Landmarks
• Indoor-6

Please run the provided scripts to prepare and download the data which has been preprocessed by running:

7scenes

./prepare_scripts/seven_scenes.sh

Cambridge Landmarks

./prepare_scripts/cambridge.sh 

Indoor-6

./prepare_scripts/indoor6.sh

Evaluation with pre-trained models

Please download the pre-trained models by running:

./prepare_scripts/download_pre_trained_models.sh

For example, to evaluate KingsCollege scene:

python runners/eval.py --dataset Cambridge --scene KingsCollege -expv pl2map

Training

python runners/train.py --dataset Cambridge --scene KingsCollege -expv pl2map_test

Supported detectors

Lines

• LSD
• DeepLSD

Points

• Superpoint

Citation

If you use this code in your project, please consider citing the following paper:

bibtexCopy
@article{bui2024pl2map,
  title={Representing 3D sparse map points and lines for camera relocalization},
  author={Bui, Bach-Thuan and Bui, Huy-Hoang and Tran, Dinh-Tuan and Lee, Joo-Ho},
  booktitle={2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year={2024}
}
@article{bui2025improved,
  title={Improved 3D Point-Line Mapping Regression for Camera Relocalization},
  author={Bui, Bach-Thuan and Bui, Huy-Hoang and Fujii, Yasuyuki and Tran, Dinh-Tuan and Lee, Joo-Ho},
  journal={arXiv preprint arXiv:2502.20814},
  year={2025}
}

This code builds on previous camera relocalization pipeline, namely D2S, please consider citing:

bibtexCopy
@article{bui2024d2s,
  title={D2S: Representing sparse descriptors and 3D coordinates for camera relocalization},
  author={Bui, Bach-Thuan and Bui, Huy-Hoang and Tran, Dinh-Tuan and Lee, Joo-Ho},
  journal={IEEE Robotics and Automation Letters},
  year={2024}
}

Acknowledgement

This code is built based on Limap, and LineTR. We thank the authors for their useful source code.