Sparse View SLAM with Ultrafast Keyframe Selector

Paper(TODO) | Video(TODO)

Sparse View SLAM with Ultrafast Keyframe Selector
Linqing Zhao*, Xiuwei Xu*, Yirui Wang*, Wenzhao Zheng, Jie Zhou, Jiwen Lu†

* Equal contribution, † Corresponding author

SV-SLAM is a SLAM system with >100fps inference speed that enables ultrafast tracking and mapping with a keyframe selector module.

News

• [2025/8/25] Code released.

Demo

SLAM System

UAV SLAM

Keyframe Selector

Mapping

Method

Method pipeline:

Getting Started

Environment Setup

conda create -n svslam python=3.11
conda activate svslam 
conda install pytorch torchvision pytorch-cuda=12.4 -c pytorch -c nvidia  # use the correct version of cuda for your system
pip install -r requirements.txt

# install lietorch
cd thirdparty
git clone --recursive https://github.com/princeton-vl/lietorch.git
cd lietorch
pip install -e .
cd ../..

Then, download MASt3R checkpoint from MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric and put in thirdparty/mast3r/checkpoints.

Download selector checkpoint from here and put it in selector.

Run on 7-Scenes and TUM-RGBD

To run with 7-Scenes or TUM-RGBD dataset, modify config/data/{7scenes,tum}.yaml according to your dataset location and sequence to use, then run

python slam.py --expdir 7scenes_demo \
    --data_config config/data/7scenes.yaml \
    --slam_config config/slam/7scenes.yaml

Run on Custom Data

To run with custom data (unposed RGB image sequence), modify config/data/simple.yaml accoding to data location and camera intrinsics. You can also modify SLAM configuration in config/slam/simple.yaml.

python slam.py --expdir demo \
    --data_config config/data/simple.yaml \
    --slam_config config/slam/simple.yaml

TODO

Acknowledgement

We base our work on the great open-sourced repo MASt3R. We also a lot for the excellent works RAFT and lietorch.

Citation

If you find this project helpful, please consider citing the following paper:

TODO