This repository contains the reference implementation of NFL-BA (Near-Field Light Bundle Adjustment).

NFL-BA is implemented as a plug-and-play photometric optimization module, and in this repo we demonstrate it on top of the MonoGS SLAM / reconstruction backbone.

The idea from this paper (correcting the image formulation model and bundle adjustment) works on multiple different backbones

• Backbone: MonoGS (monocular Gaussian splatting SLAM / reconstruction).
• Approach: Plug-and-play near-field lighting bundle adjustment:
  • We keep the overall MonoGS pipeline intact.
  • We introduce a physically motivated near-field lighting model into the renderer and loss.
• Main code changes:
  • Renderer:
    • Extensions to the rendering code to model near-field illumination (e.g., inverse-square falloff, angle-dependent terms).
  • Loss functions:
    • Modifications to incorporate the near-field photometric residuals into the optimization.
• Other changes:
  • Minor changes to the CUDA rasterizer.
  • Minor adjustments in configuration files and experiment scripts to enable NFL-BA.

The environment and dependencies are intended to be identical or very close to the original MonoGS implementation.

conda env create -f environment.yml
conda activate MonoGS

pip install -e submodules/diff-gaussian-rasterization

Use the same Python version, CUDA / PyTorch versions, and compilation steps recommended by MonoGS.

Below are example commands for reproducing experiments on C3VD and on self-captured data. Adjust paths as needed.

python slam.py \
    --config configs/mono/c3vd/base_config.yaml \
    --experiment c3vd \
    --depth c3vd \
    --scene /path/to/C3VD/cecum_t1_a_under_review

Key arguments:

• --config: Base configuration file for MonoGS + NFL-BA on C3VD.
• --experiment: Experiment tag (c3vd in our setup).
• --depth: Depth mode (c3vd uses the dataset depth maps, while pps uses depth predicted by PPSNet).
• --scene: Path to the specific C3VD sequence.

For our own self-captured sequences (e.g., pooled data from real-world scenes), we use:

python slam.py \
    --config configs/rgbd/wild/base_config.yaml \
    --experiment wild \
    --scene /path/to/data/pool_ON

Again, you should adapt:

• --config to match your configuration for self-captured RGBD or monocular data.
• --scene to the directory containing your sequence.

If you find this code useful in your research, please cite:

@inproceedings{
    beltran2025nflba,
    title={{NFL}-{BA}: Near-Field Light Bundle Adjustment for {SLAM} in Dynamic Lighting},
    author={Andrea Dunn Beltran and Daniel Rho and Marc Niethammer and Roni Sengupta},
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
    year={2025},
    url={https://openreview.net/forum?id=AmZ7uHDJiR}
}

• This codebase is built on top of MonoGS, and we are grateful to the original authors for releasing their implementation.
• Please also cite the MonoGS paper and repository if you build on this work.

• Release per-scene training / reconstruction code used in the paper.
• Release self-captured data
• Release evaluation scripts for pose and reconstruction metrics.