MBA-SLAM: Motion Blur Aware Dense Visual SLAM with Radiance Fields Representation

TPAMI 2025

Overview

Given a sequence of severe motion blurred images and depth, MBA-SLAM can accurately estimate the local camera motion trajectory of each blurred image within the exposure time and recovers the high quality 3D scene.

Pipeline

Tracking: Our motion blur-aware tracker directly estimates the camera motion trajectory during the exposure time. Mapping: Our mapper generates virtual sharp images along the camera trajectory.

Quickstart

1. Setup environment (Both setup options are still under testing and will be updated continuously.)

MBA-SLAM requires an NVIDIA GPU and CUDA. We provide two ways to set up the environment.

Option A: Conda environment

Create the Python environment from environment.yaml:

conda env create -f environment.yaml
conda activate mba-slam
pip install -r requirements.txt
pip install pypose --no-deps

Build and install the MBA-VO extension:

cd external/MBA-VO
cmake -GNinja -B build
cmake --build build -t install
cd python
pip install .
cd ../../..

If cmake --build build -t install cannot write to /usr/local, run that install step with sudo, then run sudo ldconfig. If CMake cannot find Eigen, OpenCV, or CUDA, install the corresponding system packages first, or use the Docker environment below.

Option B: Docker environment

The project Docker setup uses two images. Dockerfile.dev builds the base environment and tags it as mba-slam-base; Dockerfile then uses FROM mba-slam-base and installs MBA-VO plus the remaining Python requirements.

Build the base image:

docker build -f Dockerfile.dev -t mba-slam-base .

Build the MBA-SLAM image:

docker build -f Dockerfile -t mba-slam .

Run the container from the project root:

docker run --gpus all --ipc=host -it --rm \
  -v "$(pwd)":/workspace/MBA-SLAM \
  -w /workspace/MBA-SLAM \
  mba-slam bash

Inside the container, you can quickly check CUDA and MBA-VO:

python -c "import torch; import pymbavo; print('cuda:', torch.cuda.is_available())"

The Docker build compiles several CUDA/C++ dependencies and can take a long time. Make sure the host has Docker, the NVIDIA driver, and NVIDIA Container Toolkit installed.

2. Prepare data and config

Put the dataset under the path used by the config file. For example, configs/MbaVo/archviz1.yaml expects:

data:
  input_folder: Datasets/MBA-VO/archviz1
  output: output/MBA-VO/archviz1

3. Training

Run MBA-SLAM on the ArchViz1 example config:

python run.py ./configs/MbaVo/archviz1.yaml

Outputs are written to the config's data.output directory.

4. Evaluate ATE

After training, evaluate the absolute trajectory error with:

python ./src/tools/eval_ate.py ./configs/MbaVo/archviz1.yaml

5. Evaluate reconstruction

Evaluate 3D and 2D reconstruction metrics with:

python ./src/tools/eval_recon.py --rec_mesh <your.ply path> --gt_mesh <gt.ply path> -3d -2d

Citation

If you find this useful, please consider citing our paper:

@article{wang2025mbaslam,
      title     = {MBA-SLAM: MBA-SLAM: Motion Blur Aware Dense Visual SLAM with Radiance Fields Representation},
      author    = {Wang, Peng and Zhao, Lingzhe and Zhang, Yin and Zhao, Shiyu and Liu, Peidong},
      journal   = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
      year      = {2025},
      publisher = {IEEE}
  }

  or

@article{wang2024mbaslam,
      title     = {MBA-SLAM: Motion Blur Aware Gaussian Splatting SLAM},
      author    = {Wang, Peng and Zhao, Lingzhe and Zhang, Yin and Zhao, Shiyu and Liu, Peidong},
      journal   = {arXiv preprint arXiv:2411.08279},
      year      = {2024}
  }

Acknowledgment

We would like to thank the authors of CoSLAM, ESLAM, gsplat, and MonoGS for their excellent open-source projects and helpful discussions.