Neural Graph Mapping

This repository contains the official implementation for the paper Neural Graph Mapping for Dense SLAM with Efficient Loop Closure by Leonard Bruns, Jun Zhang, and Patric Jensfelt. By anchoring neural fields to the pose graph of a sparse visual SLAM system, we can do dense neural mapping while supporting large-scale loop closures (i.e., map deformation) without requiring full reintegration.

kintinuous.mp4

Installation

pixi

To quickly get the code running and reproduce the results you can use pixi. First, install pixi. Then run the following command to install the package, download the data, and run an example scene (the datasets will by default be stored in ./datasets/; you can modify the dataset dir in the .pixi.sh file):

pixi run nrgbd_br --rerun_vis True

To run all the scenes and datasets you can run:

pixi run all

You can optionally add arguments to all commands by settings NGM_EXTRA_ARGS prior to running the command. For example, to run all datasets with visualization enabled run:

NGM_EXTRA_ARGS="--rerun_vis True --rerun_save True" pixi run all

Manual

First you need to install torch==2.2.* and the corresponding CUDA version (such that nvcc is available and matches the torch's CUDA version). This is necessary because some dependencies are installed from source.

To install this package and all dependencies, clone this repo, and run

pip install --no-build-isolation -e .

Development

• Use pip install --no-build-isolation -e . to install the package in editable mode
• Use pip install -r requirements-dev.txt to install dev tools

Reference

If you compare with our method or find this code useful in your research, consider citing our preprint:

@article{bruns2024neural,
  title={Neural Graph Mapping for Dense SLAM with Efficient Loop Closure},
  author={Bruns, Leonard and Zhang, Jun and Jensfelt, Patric},
  journal={arXiv preprint arXiv:2405.03633},
  year={2024}
}