SLAM3R for macOS (Apple Silicon)

This repository contains a modified version of the original SLAM3R project, adapted to run on Apple Silicon (M1/M2/M3) Macs using Metal Performance Shaders (MPS) for GPU acceleration instead of NVIDIA CUDA.

---

Acknowledgments & Credit

This project would not be possible without the incredible work of the original authors. This repository is a direct fork and adaptation of their research and code. Please be sure to cite their original work.

• Primary Source: PKU-VCL-3DV/SLAM3R
  • Paper: SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos by Yuzheng Liu, Siyan Dong, et al.
• Influential Work: The authors also credit cvg/nicer-slam.

This fork is intended solely to make this powerful tool accessible to users on Apple hardware.

---

macOS Adaptation

The original SLAM3R codebase contains hardcoded calls to the NVIDIA CUDA API, which is not available on macOS. The following key modifications were made to enable compatibility:

1. Modified recon.py:

   • Removed hardcoded CUDA device selection (torch.cuda.set_device).
   • Added logic to auto-detect the available hardware, prioritizing mps (Apple GPU), then cuda, and falling back to cpu.
   • The selected device is passed as an argument to the model loading and pipeline functions.

2. Modified slam3r/utils/device.py:

   • The MyNvtxRange class, used for profiling, was wrapped in a conditional check (if torch.cuda.is_available():) to prevent CUDA-specific calls on non-NVIDIA hardware.
   • Helper functions that were accidentally removed in previous edits have been restored to ensure the script runs correctly.

These changes make the codebase hardware-agnostic, allowing it to run on both the original NVIDIA platform and on Apple Silicon Macs.

---

Installation on macOS

This guide assumes you have Homebrew and Python 3.11 installed.

1. Clone the Repository:

   git clone [https://github.com/jadouse5/SLAM3R_MAC]
   cd SLAM3R_MAC 
   

2. Create a Python Virtual Environment:

   python3.11 -m venv myenv
   source myenv/bin/activate
   

3. Install Dependencies in Order: The installation order is important to ensure compatibility.

   # Upgrade pip
   pip install --upgrade pip
   
   # Install the correct PyTorch version for Apple Silicon
   pip install torch torchvision torchaudio
   
   # Install the remaining packages from requirements.txt
   pip install -r requirements.txt
   

---

Usage

This repository includes a Gradio web interface for easy use.

1. Launch the App: Make sure your (myenv) is active and you are in the main project directory.

   python app.py
   

2. Use the Interface: Open the provided URL in your web browser. You can upload a video to generate a new 3D point cloud (.ply) file or upload an existing .ply file to view it.

Citations

@article{slam3r, title={SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos}, author={Liu, Yuzheng and Dong, Siyan and Wang, Shuzhe and Yin, Yingda and Yang, Yanchao and Fan, Qingnan and Chen, Baoquan}, journal={arXiv preprint arXiv:2412.09401}, year={2024} }