MISO: Multiresolution Submap Optimization for Efficient Globally Consistent Neural Implicit Reconstruction

📄 Paper | 🌐 Project Website

This is the code repository for MISO, a hierarchical optimization approach that leverages multiresolution submaps to achieve efficient and scalable neural implicit reconstruction. For local SLAM within each submap, we develop a hierarchical optimization scheme with learned initialization that substantially reduces the time needed to optimize the implicit submap features. To correct estimation drift globally, we develop a hierarchical method to align and fuse the multiresolution submaps, leading to substantial acceleration by avoiding the need to decode the full scene geometry.

Outline

0. Set up

1. Mapping with groundtruth robot poses

2. Submap alignment

3. Full SLAM

Set up

A. Download necessary data

Create a data folder under the project root,

mkdir data && cd data

Download the ScanNet and Newer College (optional) datasets from the google drive, and unzip it to the data folder.

Further make a results folder under the project root,

cd .. && mkdir results && cd results

Download the pretrained decoders from the google drive and unzip it to the results folder

/root/of/the/project
    |———data
    |     |———ScanNet
    |     |      |———scans
    |     |———Newer_College
    |            |———ply
    |            |———...
    |———results
    |     |———trained_decoders
    |            |———decoder_indoor.pt
    |            |———decoder_quad.pt
    |——— ....
    |    

B. Create the environment

We offer two ways of setting up the environment: docker and conda.

B1. Use the prebuilt docker image

docker pull hwcao17/miso_lab

xhost +local:

docker run -it --mount type=bind,source=/path/to//MISO-DCIST-Lab,target=/home/MISO-DCIST-Lab --gpus all -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix hwcao17/miso_lab:latest

B2. Create a Conda environment

Then, let's prepare the python environment, run the following command

conda env create -f environment.yaml

Note that the torch installation could be tricky. I have CUDA 11.6 and use the following command

pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116

You may also find all the installation commands in install.sh.

Finally, install the grid opt package

pip install -e .

Mapping with groundtruth robot poses

In this example, we use the multiresolution feature grid to build individual submaps on the ScanNet dataset. Run the following command:

python demo/build_submaps.py

Take Submap 0 for example, we will see the coarse-level and fine level mesh constructions as following:

Submap alignment

In this example, we run the proposed latent space alignment method to register all submaps consistently in the global reference frame. The following visualizes the alignment process as the submap poses are optimized.

To only use the coarse-level features to align the submaps

python demo/align_submaps.py --feature_levels 0

To use both the coarse-level and the fine-level features to align the submaps

python demo/align_submaps.py --feature_levels 0 1

To use only the SDF prediction to align the submaps

python demo/align_submaps.py --feature_levels --use_sdf

Run the full SLAM

To run on the outdoor quad scene from the Newer College dataset:

python demo/full_slam_newer_college.py

Citation

If you found this work useful, we would appreciate if you could cite our work:

• Y. Tian, H. Cao, S. Kim, N. Atanasov. MISO: Multiresolution Submap Optimization for Efficient Globally Consistent Neural Implicit Reconstruction. arXiv:2504.19104.

@inproceedings{tian2025miso,
 title={{MISO}: Multiresolution Submap Optimization for Efficient Globally Consistent Neural Implicit Reconstruction},
 author={Tian, Yulun and Cao, Hanwen and Kim, Sunghwan and Atanasov, Nikolay},
 booktitle={Robotics: Science and Systems (RSS)},
 year={2025}
}

Acknowledgments

This repo was built using some awesome third party functions from iSDF, PIN-SLAM, nglod, and cuda-gridsample-grad2.

We gratefully acknowledge support from ARL DCIST CRA W911NF-17-2-0181, ONR N00014-23-1-2353, and NSF CCF-2112665 (TILOS).

License

BSD License