CVI2-UniLU/SPARK-2026

Dataset Card for SPARK 2026

SPARK 2026 is a dataset of spacecraft imagery with bounding box labels and segmentation masks, released as part of the SPARK 2026 Challenge for Multi-Task Spacecraft Perception. It covers 10 spacecraft classes rendered under diverse space-like conditions. The intent is to design a single, powerful model capable of performing spacecraft classification, detection, and fine-grained segmentation of spacecraft components—regardless of spacecraft type. The focus is on efficiency and performance, encouraging the development of compact, high-performing models suitable for deployment on resource-constrained space platforms.

Dataset Details

Dataset Description

The dataset provides RGB images and corresponding bounding box labels, segmentation masks for 10 spacecraft targets, split into training and validation sets per object. It is designed to benchmark spacecraft recognition and semantic segmentation methods for Space Situational Awareness.

• Curated by: SnT, University of Luxembourg (CVI² group)
• License: CC-BY-4.0

Dataset Sources

• Challenge: SPARK 2026 Challenge (Stream 1)
• Related release: SPARK 2024 on Zenodo

Uses

Direct Use

• Spacecraft semantic segmentation
• Spacecraft classification / recognition
• Benchmarking vision models for space-based applications (e.g., SSA, inspection)

Out-of-Scope Use

The dataset is not intended for tasks unrelated to spaceborne vision, and models trained on it should not be assumed to generalize to real flight imagery without appropriate domain validation.

Dataset Structure

Create a data folder and download the training and validation archives into it. After unzipping the *.zip archives, the tree structure of data must follow:

data/
├── images/
│   ├── object_1/
│   │   ├── train/
│   │   │   └── img1...
│   │   └── val/
│   │       └── img1...
│   └── object_2/
│       ├── train/
│       │   └── img1...
│       └── val/
│           └── img1...
├── mask/
│   ├── object_1/
│   │   ├── train/
│   │   │   └── mask1...
│   │   └── val/
│   │       └── mask1...
│   └── object_2/
│       ├── train/
│       │   └── mask1...
│       └── val/
│           └── mask1...
├── train.csv
└── val.csv

The visualize_data.ipynb notebook contains basic functions to load and display dataset samples.

Class Labels

Class name	Index
VenusExpress	0
Cheops	1
LisaPathfinder	2
ObservationSat1	3
Proba2	4
Proba3	5
Proba3ocs	6
Smart1	7
Soho	8
XMM Newton	9

Dataset Creation

Curation Rationale

The dataset was created to support the SPARK Challenge, an initiative promoting the development and benchmarking of robust vision algorithms for spacecraft recognition and segmentation under realistic, space-representative imaging conditions.

Source Data

Data Collection and Processing

Images were generated with a photorealistic rendering pipeline of spacecraft models under varied poses, distances, and illumination conditions, with pixel-accurate segmentation masks produced alongside each image.

Bias, Risks, and Limitations

The imagery is synthetic; models trained solely on this data may exhibit a sim-to-real performance gap when deployed on real orbital imagery. Users should validate on real or mission-representative data before operational use.

Citation

BibTeX:

@dataset{rathinam_2024_10908215,
  author       = {Rathinam, Arunkumar and
                  Mohamed Ali, Mohamed Adel and
                  Gaudilliere, Vincent and
                  Aouada, Djamila},
  title        = {SPARK 2024: Datasets for Spacecraft Semantic
                  Segmentation and Spacecraft Trajectory Estimation},
  month        = feb,
  year         = 2024,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.10908215},
  url          = {https://doi.org/10.5281/zenodo.10908215}
}

APA:

Rathinam, A., Mohamed Ali, M. A., Gaudilliere, V., & Aouada, D. (2024). SPARK 2024: Datasets for Spacecraft Semantic Segmentation and Spacecraft Trajectory Estimation [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10908215

Dataset Card Contact

spark@uni.lu CVI², SnT, University of Luxembourg