SLAM&Render: A Benchmark for the Intersection Between Neural Rendering, Gaussian Splatting and SLAM

Samuel Cerezo · Gaetano Meli · Tomas Berriel Martins · Kirill Safronov · Javier Civera

2025

ArXiv | Video | Project Page

---

🎬 SLAM&Render: Video Overview

Explore the SLAM&Render dataset with this short video showcasing the five key sequences included in the project. Click below to watch the video!

Watch the full introduction to the SLAM&Render dataset and see how the dataset integrates neural rendering, Gaussian splatting, and SLAM in action.

---

📌 Overview

🚨 Current Datasets' Limitations: Existing datasets fail to address critical challenges in the following fields:

• 🧩 Multi-modality and Sequentiality in SLAM.
• 🌍 Generalization across Viewpoints and Lighting Conditions in Neural Rendering.

---

✨ Introducing SLAM&Render: A novel dataset designed to explore the intersection of both SLAM and Neural Rendering, covering:

• 📊 40 Real-World Sequences.
• 🔄 Synchronized RGB, Depth, IMU, Robot Kinematics, and Ground-Truth Poses.

---

📁 Dataset Structure

Each sequence contains:

sequence_name/
├── rgb/                          # RGB images (30 Hz)
├── depth/                        # Aligned depth images
├── robot_data/joint_states.txt   # Robot joint encoders (25 Hz)
├── robot_data/flange_poses.txt   # Forward kinematics pose (25 Hz)
├── associations.txt              # Association between images
├── imu.txt                       # Accelerometer + gyroscope (210 Hz)
├── groundtruth_raw.csv           # Raw MoCap ground truth (120 Hz)
├── groundtruth.txt               # MoCap ground truth with Unix epoch timestamp(120 Hz)

See data/README.md for full details.

---

🚀 Getting Started

git clone https://github.com/samuel-cerezo/slam-render.git
cd slam-render
pip install -r requirements.txt

To download a sequence:

python scripts/download_data.py --sequence 4-natural-train

---

🛠️ Examples & Tools

We provide utility scripts to align and use the dataset easily:

Script	Description
scripts/temporal_align.py	Align timestamps between camera and gt data.
scripts/fFlange2world.py	Align poses to a world frame using motion capture.
scripts/ROSBAG2TUM.py	Convert .bag file into a TUM format.
scripts/reading_mocap_CSV.py	Convert mocap raw data into a txt with the corresponding unix epoc timestamp.
scripts/download_data.py	For downloading dataset sequences
slamrender/alignment_utils.py	Core functions for spatial and temporal alignment.

See the notebook notebooks/example_usage.ipynb for a full pipeline.

---

🧪📊 Evaluation

In this section, we provide the evaluation results of the SLAM&Render dataset. The following figures show the key metrics for evaluating the performance of models trained on this dataset:

1. RMSE ATE Evaluation The following figure illustrates the RMSE ATE results obtained for setup-4 under cold conditions. The trajectory result using MonoGS is shown in a). Similarly when kinematic data as initial seed is used, the result is shown in b).

2. Novel View Synthesis Evaluation

We also assess the performance of NVS on the dataset. To highlight the usefulness of the independent test camera trajectories, we evaluated the performance of two state-of-the-art baselines: Gaussian Splatting; and FeatSplat. The following image compares the rendered output to the ground truth.

---

🤝 Acknowledgements

This work includes many open-source projects:

• MonoGS (Gaussian Splatting SLAM)

---

📄 License

Released under the LICENSE.md.

---

📚 Citation

If you find this dataset useful, please cite us:

@misc{cerezo_slamrender,
      title={SLAM&Render: A Benchmark for the Intersection Between Neural Rendering, Gaussian Splatting and SLAM}, 
      author={Samuel Cerezo and Gaetano Meli and Tomás Berriel Martins and Kirill Safronov and Javier Civera},
      year={2025},
      eprint={2504.13713},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2504.13713}, 
}

---

🌐 Links

• 📄 Project Page

---