CVI2-UniLU/SPARK-2022

SPARK 2022 — Stream 1 (Spacecraft Detection)

Stream 1 of the SPARK 2022 dataset (SPAcecraft Recognition leveraging Knowledge of the space environment): space-borne imagery of 10 spacecraft plus a debris class, for object detection and classification. Each image contains exactly one target annotated with a single bounding box and class label.

Dataset summary

Images	110,000 JPEG, 1024 × 1024, RGB
Annotations	1 bounding box + class per image
Classes	11 (10 spacecraft + debris)
Splits	train 66,000 · val 22,000 · test 22,000
Class balance	fully balanced (6,000 / 2,000 / 2,000 per class per split)

Dataset structure

├── labels/
│   ├── train.csv
│   ├── val.csv
│   └── test.csv
├── train/          # 66,000 .jpg images
├── val/            # 22,000 .jpg images
├── test/           # 22,000 .jpg images
└── visualize_labels.py

Label format

Each CSV has the header filename,class,bbox:

filename,class,bbox
img057676.jpg,lisa_pathfinder,"[633, 120, 789, 279]"

• filename matches the image file in the corresponding split folder.
• bbox is [xmin, ymin, xmax, ymax] in absolute pixel coordinates, with the origin at the top-left corner of the image (x = column, y = row).
• class is the class name; see the index mapping below.

Classes

Class	Name	Index
Proba 2	proba_2	0
Cheops	cheops	1
Debris	debris	2
Double star	double_star	3
Earth Observation Sat 1	earth_observation_sat_1	4
Lisa Pathfinder	lisa_pathfinder	5
Proba 3 CSC	proba_3_csc	6
Proba 3 OCS	proba_3_ocs	7
Smart 1	smart_1	8
Soho	soho	9
Xmm Newton	xmm_newton	10

Usage

import ast
from pathlib import Path

import pandas as pd
from PIL import Image

CLASS_TO_INDEX = {
    "proba_2": 0, "cheops": 1, "debris": 2, "double_star": 3,
    "earth_observation_sat_1": 4, "lisa_pathfinder": 5, "proba_3_csc": 6,
    "proba_3_ocs": 7, "smart_1": 8, "soho": 9, "xmm_newton": 10,
}

root = Path(".")
split = "train"

df = pd.read_csv(root / "labels" / f"{split}.csv")
df["bbox"] = df["bbox"].apply(ast.literal_eval)          # [xmin, ymin, xmax, ymax]
df["label"] = df["class"].map(CLASS_TO_INDEX)

row = df.iloc[0]
image = Image.open(root / split / row["filename"])
xmin, ymin, xmax, ymax = row["bbox"]

For a PyTorch object-detection pipeline (e.g. torchvision), targets follow directly since the boxes are already in xyxy format:

import torch

target = {
    "boxes": torch.tensor([row["bbox"]], dtype=torch.float32),   # (1, 4) xyxy
    "labels": torch.tensor([row["label"]], dtype=torch.int64),
}

Visual inspection

The bundled script plots random or specific samples with their boxes drawn:

python3 visualize_labels.py val --num 6 --seed 42
python3 visualize_labels.py train --class debris
python3 visualize_labels.py test --filenames img057676.jpg

Dataset origin

The SPARK dataset was created by the CVI² group at SnT, University of Luxembourg, for the SPARK challenge on spacecraft detection and recognition.

Citation

If you use this dataset, please use the citation below:

@dataset{rathinam_2022,
  author       = {Rathinam, Arunkumar and
                  Gaudilliere, Vincent and
                  Mohamed Ali, Mohamed Adel and
                  Ortiz Del Castillo, Miguel and
                  Pauly, Leo and
                  Aouada, Djamila},
  title        = {SPARK 2022 Dataset : Spacecraft Detection and
                   Trajectory Estimation},
  month        = jun,
  year         = 2022,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.6599762},
  url          = {https://doi.org/10.5281/zenodo.6599762},
}