Pseudo Depth Meets Gaussian: A Feed-forward RGB SLAM Baseline

Paper

Pseudo Depth Meets Gaussian: A Feed-forward RGB SLAM Baseline
Linqing Zhao*, Xiuwei Xu*, Yirui Wang, Hao Wang, Wenzhao Zheng, Yansong Tang, Haibin Yan†, Jiwen Lu

* Equal contribution, † Corresponding author

DepthGS is a RGB-SLAM system that integrates monocular depth estimation, 3DGS scene representation and a neural network-based pose estimating module.

News

• [2025/06/16] Paper accepted to IROS 2025 🎊
• [2025/08/01] Code released 🚀

Demo

Method

Method pipeline:

Getting Started

Setup Environment

Our code has been developed with python 3.10, CUDA 11.7 and torch 1.13.1. It has been tested with python 3.10, CUDA 12.4 and torch 2.5.1, which also works.

1. Create environment & install necessary packages

conda create -n depthgs python=3.10
conda activate depthgs
conda install pytorch torchvision pytorch-cuda=12.4 -c pytorch -c nvidia
pip install -r requirements.txt

Then, install torch-scatter. Installing via pip may cause CUDA version incompatability (.../torch_scatter/_version_cuda.so: undefined symbol). In this case, download and install the appropriate version from here (you shall navigate to find the correct python, torch and CUDA version). For example,

wget https://data.pyg.org/whl/torch-2.5.0%2Bcu124/torch_scatter-2.1.2%2Bpt25cu124-cp310-cp310-linux_x86_64.whl
pip install torch_scatter-2.1.2+pt25cu124-cp310-cp310-linux_x86_64.whl

2. Install 3DGS rasterization engine:

git clone git@github.com:JonathonLuiten/diff-gaussian-rasterization-w-depth.git
cd diff-gaussian-rasterization-w-depth
python setup.py install
pip install -e .

3. Install lietorch, pytorch-scatter and backends of DROID-SLAM:

git clone --recursive https://github.com/princeton-vl/DROID-SLAM.git
cd DROID-SLAM
# this will take a while
pip install thirdparty/lietorch
pip install thirdparty/pytorch_scatter
pip install -e .

Download droid.pth to DROID-SLAM/droid.pth.

4. Install monocular depth estimator e.g. UniDepthV2

Prepare Dataset

Run your MDE model to get (H, W) depth maps, and save them as .npy files. Arrange data for each sequence as:

your_data_dir/seq0
├── intrinsics.txt
├── color
│   ├── 000000.png
│   ├── 000001.png
│   ├── ...
└── depth
    ├── 000000.npy
    ├── 000001.npy
    ├── ...

Where intrinsics.txt contains:

fx fy cx cy

Then, modify SplaTAM/configs/custom/splatam.py according to your data directory. (See NOTE in the script.)

Run DepthGS system

python slam.py SplaTAM/configs/custom/splatam.py

Results will be saved to experiments/custom/slam_run. You can set the output directory by modifying SplaTAM/configs/custom/splatam.py.

Then, run

python SplaTAM/scripts/export_ply.py SplaTAM/configs/custom/splatam.py

to get splat.ply, which can be visualized with tools such as SuperSplat.

Acknowledgement

We base our work on the great work SplaTAM. We also a lot for the excellent work DROID-SLAM.

Citation

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

@article{zhao2025pseudo, 
      title={Pseudo Depth Meets Gaussian: A Feed-forward RGB SLAM Baseline}, 
      author={Linqing Zhao and Xiuwei Xu and Yirui Wang and Hao Wang and Wenzhao Zheng and Yansong Tang and Haibin Yan and Jiwen Lu},
      journal={arXiv preprint arXiv:2508.04597},
      year={2025}
}