pretty_name: GS2E
paperswithcode_id: robust-e-nerf-synthetic-event-dataset
license: cc-by-4.0
viewer: false
size_categories:
- 1K<n<10K
📦 GS2E: Gaussian Splatting is an Effective DataGenerator for Event Stream Generation
Submission to NeurIPS 2025 D&B Track, Under Review.
🧾 Dataset Summary
GS2E (Gaussian Splatting for Event stream Extraction) is a synthetic multi-view event dataset designed to support high-fidelity 3D scene understanding, novel view synthesis, and event-based neural rendering. Unlike previous video-driven or graphics-only event datasets, GS2E leverages 3D Gaussian Splatting (3DGS) to generate geometry-consistent photorealistic RGB frames from sparse camera poses, followed by physically-informed event simulation with adaptive contrast threshold modeling. The dataset enables scalable, controllable, and sensor-faithful generation of realistic event streams with aligned RGB and camera pose data.
📚 Dataset Description
Event cameras offer unique advantages—such as low latency, high temporal resolution, and high dynamic range—making them ideal for 3D reconstruction and SLAM under rapid motion and challenging lighting. However, the lack of large-scale, geometry-consistent event datasets has hindered the development of event-driven or hybrid RGB-event methods.
GS2E addresses this gap by synthesizing event data from sparse, static RGB images. Using 3D Gaussian Splatting (3DGS), we reconstruct high-fidelity 3D scenes and generate dense camera trajectories to render blur-free and motion-blurred sequences. These sequences are then processed by a physically-grounded event simulator, incorporating adaptive contrast thresholds that vary across scenes and motion profiles.
The dataset includes:
- 21 multi-view event sequences across 7 scenes and 3 difficulty levels (easy/medium/hard)
- Per-frame photorealistic RGB renderings (clean and motion-blurred)
- Ground truth camera poses
- Geometry-consistent synthetic event streams
- Consistent intrinsics and camera paths with NeRF Synthetic Dataset
The result is a simulation-friendly yet physically-informed dataset for training and evaluating event-based 3D reconstruction, localization, SLAM, and novel view synthesis.
If you use this synthetic event dataset for your work, please cite:
TBD
Dataset Structure and Contents
This synthetic event dataset is organized first by scene, then by level of difficulty. Each sequence recording is given in the form of a ROS bag named esim.bag, with the following data streams:
| ROS Topic | Data | Publishing Rate (Hz) |
|---|---|---|
/cam0/events |
Events | - |
/cam0/pose |
Camera Pose | 1000 |
/imu |
IMU measurements with simulated noise | 1000 |
/cam0/image_raw |
RGB image | 250 |
/cam0/depthmap |
Depth map | 10 |
/cam0/optic_flow |
Optical flow map | 10 |
/cam0/camera_info |
Camera intrinsic and lens distortion parameters | 10 |
It is obtained by running the improved ESIM with the associated esim.conf configuration file, which references camera intrinsics configuration files pinhole_mono_nodistort_f={1111, 1250}.yaml and camera trajectory CSV files {hemisphere, sphere}_spiral-rev=4[...].csv.
The validation and test views of each scene are given in the views/ folder, which is structured according to the NeRF synthetic dataset (except for the depth and normal maps). These views are rendered from the scene Blend-files, given in the scenes/ folder. Specifically, we create a Conda environment with Blender as a Python module installed, according to these instructions, to run the bpy_render_views.py Python script for rendering the evaluation views.
Setup
- Install Git LFS according to the official instructions.
- Setup Git LFS for your user account with:
git lfs install - Clone this dataset repository into the desired destination directory with:
git lfs clone https://huggingface.co/datasets/Falcary/GS2E - To minimize disk usage, remove the
.git/folder. However, this would complicate the pulling of changes in this upstream dataset repository.