This repository contains the training code for SLAM-Former.

Please refer to the main README for environment setup and data download instructions.

After downloading, update the dataset root paths in config/mytrain.yaml

Also set the pretrained checkpoint path:

pretrained: /your/path/to/pi3.pth

We use Accelerate for distributed training. Configure your setup with:

accelerate config

cd src
accelerate launch --num_processes=1 mytrain.py

cd src
accelerate launch --num_processes=8 --multi_gpu mytrain.py

This codebase is built on top of the following excellent projects:

• DUSt3R — the foundational 3D reconstruction framework we build upon
• StreamVGGT — streaming architecture and training utilities
• VGGT — geometry and pose encoding

We sincerely thank the authors of these works for open-sourcing their code.

If you find this work useful, please cite:

@article{slam-former,
    title={SLAM-Former: Putting SLAM into One Transformer},
    author={Yijun Yuan, Zhuoguang Chen, Kenan Li, Weibang Wang, and Hang Zhao},
    journal={arXiv preprint arXiv:2509.16909},
    year={2025}
}