E-SPARK: Event-based SPARK2021 Dataset for Space Object Detection

Introduction

This public dataset accompanies the paper: “M²Former: Enhancing Event-Based RT-DETR for Robust and Lightweight Space Object Detection”. More details can be found on the project page.

It consists of two components:

E-SPARK: A large-scale synthetic event dataset for detector training and validation

Testbed: A real-world dataset collected by a ground-based hardware platform under different illumination conditions for sim2real evaluation

1. E-SPARK: Synthetic Event Data

1.1 Download SPARK2021 dataset

E-SPARK is constructed as the event-based counterpart to the SPARK2021 dataset, which comprises 150,000 high-resolution (1024×1024) photo-realistic RGB images generated by the Unity engine. It simulates diverse orbital backgrounds, sensor configurations, and lighting conditions. Object categories are listed below:

10 Satellites: AcrimSat, Aquarius, Aura, Calipso, CloudSat, CubeSat, Jason, Sentinel-6, Terra, TRMM

1 Debris Class: Five debris types grouped into a single category

Each satellite class has 12,500 samples; the Debris class contains 25,000. The dataset is split into training, validation, and test sets in a 3:1:1 ratio. Please make sure to download the dataset before following the next step.

1.2 Event Generation Pipeline

The event generation process consists of three sequential steps, transforming the original SPARK2021 RGB images into labeled event streams and dense representations. The resolution is set to 640×480.

Step 1: Virtual environment setup (v2e)

conda create -n v2e python=3.10conda activate v2efollow the v2e project...pip install -r requirements.txt

Step 2: Affine transformation and event Simulation

Script: python event_simulator.py

Input: SPARK2021 RGB images

Output: Raw event streams

Transformation: Random scale, rotation, and translation applied to each image

Step 3: Event stream reformatting and labeling

Script: python espark_preprocess.py

Input: Step 2 output and SPARK2021 labels

Output: Reformatted events.txt, YOLO-format labels, COCO-format JSONs

Step 4: Convert event streams to dense representations

Script: python espark_hdf5.py

Input: Step 3 output

Output: H5 files with three representations

Optional: RGB images reformatting

Script: python spark2021_preprocess.py

Input: SPARK2021 RGB images and labels

Output: Reformatted RGB images, YOLO-format labels, COCO-format JSONs

1.3 Directory Structure

E-SPARK
├── simulator
└── dvs640
└── splits (train, validate, test)
└── categories (11 classes)
├── events
└── dvs640
├── raw (*.txt)
│ └── splits (train, validate, test)
    ├── hdf5 (*.h5)
│ └── splits (train, validate, test)
    ├── labels (YOLO format)
│ └── splits (train, validate, test)
    └── annotations (COCO format)
│ (optional)
├── images (SPARK2021 reformatting)
├── labels (YOLO format)
└── annotations (COCO format)

2. Testbed: Real Event Data

To evaluate performance under real-world conditions, we built a ground-based testbed and collected data.

2.1 Testbed Environment

Installed in a blackout-enclosed darkroom

Equipped with three ambient and one observation light source

Relative motion emulated using two mechanical arms on linear guide rails

2.2 Equipment Setups

Sensor: DAVIS346 event camera with resolution of 346×260

Targets: CubeSat, Gaofen-13, Hubble Space Telescope

Illuminations: Normal, Overexposed, Underexposed

Samples: 600 event-RGB pairs

This is a category-agnostic detection task focused on localization. All samples are labeled as a single class, i.e., satellite.

2.3 Directory Structure

Testbed
├── normal
├── events
├── images
└── labels
├── overexposed
├── events
├── images
└── labels
└── underexposed
├── events
├── images
└── labels

Each subfolder contains 200 samples:

events/: Raw event data in format of (timestamp, x, y, polarity)

images/: Corresponding RGB images, aligned with event data

labels/: Ground-truth bounding boxes in YOLO format (class center_x center_y width height)

Citations

If you utilize this work in your research, please cite:

Our paper:

@ARTICLE{11263950,  author={Pan, Ruitao and Wang, Chenxi and Han, Bin and Zhang, Xinyu and Zhai, Zhi and Liu, Jinxin and Liu, Naijin and Chen, Xuefeng},  journal={IEEE Transactions on Geoscience and Remote Sensing},   title={M2Former: Enhancing Event-Based RT-DETR for Robust and Lightweight Space Object Detection},   year={2025},  volume={63},  pages={1-16},  keywords={Space vehicles;Cameras;YOLO;Event detection;Transformers;Data augmentation;Computer architecture;Computational modeling;Training;Lighting;Event-based vision;multiscale MetaFormer design;real-time detection Transformer (RT-DETR);space object detection},  doi={10.1109/TGRS.2025.3636122}}

SPARK2021:

@inproceedings{musallam2021spacecraft,  title={Spacecraft recognition leveraging knowledge of space environment: Simulator, dataset, competition design and analysis},  author={Musallam, Mohamed Adel and Gaudilliere, Vincent and Ghorbel, Enjie and Al Ismaeil, Kassem and Perez, Marcos Damian and Poucet, Michel and Aouada, Djamila},  booktitle={2021 IEEE International Conference on Image Processing Challenges (ICIPC)},  pages={11--15},  year={2021},  organization={IEEE}}

V2e:

@inproceedings{hu2021v2e,  title={v2e: From video frames to realistic DVS events},  author={Hu, Yuhuang and Liu, Shih-Chii and Delbruck, Tobi},  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},  pages={1312--1321},  year={2021}}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.