Gaussian-SLAM on Jetson Orin: TUM RGB-D Dataset Experiments

Hüseyin Mert Çalışkan

Graduation Project - Based on Gaussian-SLAM by Vladimir Yugay et al.

📄 Research Paper

🚀 Overview

This repository contains an adaptation of Gaussian-SLAM optimized for NVIDIA Jetson Orin platform with comprehensive experiments on the TUM RGB-D dataset. This graduation project includes memory optimizations, CUDA 12.2 compatibility, and performance benchmarks for real-time inference on edge computing hardware.

⚙️ Setting Things Up

Clone the repo:

git clone https://github.com/Kovakafa/Gaussian-SLAM.git
cd Gaussian-SLAM

Make sure that gcc and g++ paths on your system are exported:

export CC=<gcc path>
export CXX=<g++ path>

Environment Setup for Jetson Orin with CUDA 12.2

Important: The original environment file has been modified for CUDA 12.2 compatibility and Jetson Orin optimization.

# Create environment from modified conda file
conda env create -f environment_jetson.yml
conda activate gslam-jetson

Key Environment Modifications

• CUDA: 12.2 (instead of original version)
• PyTorch: 2.1.0 compatible with CUDA 12.2
• Torchvision: 0.18.0 (optimized for Jetson Orin)
• Memory optimizations: Reduced package versions for Jetson constraints

# Install additional Jetson-specific dependencies
pip install -r requirements_jetson.txt

We tested our code on Jetson Orin NX 16GB with Ubuntu 20.04 and CUDA 12.2.

🔨 Running Gaussian-SLAM on Jetson Orin

Downloading TUM RGB-D Dataset

Download the TUM RGB-D dataset:

# Create datasets directory
mkdir datasets && cd datasets

# Download TUM RGB-D sequences
wget https://vision.in.tum.de/rgbd/dataset/freiburg1/rgbd_dataset_freiburg1_xyz.tgz
wget https://vision.in.tum.de/rgbd/dataset/freiburg2/rgbd_dataset_freiburg2_xyz.tgz
wget https://vision.in.tum.de/rgbd/dataset/freiburg3/rgbd_dataset_freiburg3_long_office_household.tgz

# Extract datasets
tar -xzf rgbd_dataset_freiburg1_xyz.tgz
tar -xzf rgbd_dataset_freiburg2_xyz.tgz
tar -xzf rgbd_dataset_freiburg3_long_office_household.tgz

Or download from the official TUM RGB-D Dataset page: https://vision.in.tum.de/data/datasets/rgbd-dataset

The dataset should be organized as:

datasets/
├── rgbd_dataset_freiburg1_xyz/
├── rgbd_dataset_freiburg2_xyz/
└── rgbd_dataset_freiburg3_long_office_household/

Running the code

Start the system with the command:

python run_slam.py configs/TUM_RGBD/<config_name> --input_path <path_to_scene> --output_path <output_path>

For example:

python run_slam.py configs/TUM_RGBD/freiburg1_xyz.yaml --input_path datasets/rgbd_dataset_freiburg1_xyz --output_path output/TUM_RGBD/freiburg1_xyz

For real-time inference on Jetson Orin:

python real_time_slam.py --camera_id 0 --config configs/jetson_real_time.yaml

Jetson Orin Optimizations

Key optimizations implemented for Jetson Orin with CUDA 12.2:

Environment Optimizations:

• Modified conda environment for CUDA 12.2 compatibility
• Torchvision 0.18.0 for improved Jetson performance
• Memory-efficient package versions

Runtime Optimizations:

• Memory Management: Adaptive Gaussian pruning for 16GB memory constraint
• Mixed Precision: FP16 inference with CUDA 12.2 optimizations
• Batch Processing: Optimized batch sizes for Jetson Orin architecture
• Real-time Pipeline: Asynchronous processing and adaptive frame rate control

Performance Results on TUM RGB-D

Performance metrics on Jetson Orin NX 16GB with TUM RGB-D dataset:

Sequence	RMSE (cm)	FPS	Memory Usage (GB)	Power (W)
freiburg1_xyz	2.14	12.3	14.2	22.5
freiburg2_xyz	1.89	11.8	13.8	23.1
freiburg3_long	2.31	10.2	15.1	24.3

Comparison with Desktop GPU:

• Jetson Orin NX: 11.4 fps avg, 23.3 W power consumption
• RTX 4090: 38.7 fps avg, 450+ W power consumption
• Power efficiency: ~30% of desktop performance at ~5% power consumption

Detailed results and analysis available in our research paper.

Environment File Modifications

The original environment.yml has been modified for Jetson Orin compatibility:

Changed Dependencies:

# Original vs Modified
pytorch: 1.13.0 → 2.1.0+cu122
torchvision: 0.14.0 → 0.18.0
cudatoolkit: 11.7 → 12.2

Added Jetson-specific packages:

- jetson-stats
- jtop
- tensorrt  # for inference optimization

View complete environment file: environment_jetson.yml

📄 Research Paper

Our detailed research paper covering the Jetson Orin adaptation, optimization strategies, and TUM RGB-D experiments:

📋 GS-SLAM_JetsonOrin.pdf

📌 Citation

If you find our adaptation useful, please cite both the original work and our contribution:

Original Gaussian-SLAM:

@misc{yugay2023gaussianslam,
      title={Gaussian-SLAM: Photo-realistic Dense SLAM with Gaussian Splatting}, 
      author={Vladimir Yugay and Yue Li and Theo Gevers and Martin R. Oswald},
      year={2023},
      eprint={2312.10070},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

This work:

@misc{caliskan2024gaussianslam,
      title={Gaussian-SLAM on Jetson Orin: Real-time Dense SLAM for Edge Computing with TUM RGB-D Dataset}, 
      author={Hüseyin Mert Çalışkan},
      year={2024},
      note={Graduation Project - Available at: https://github.com/Kovakafa/Gaussian-SLAM}
}