SPARK — training weights
Companion weights for the SPARK spacecraft-perception experiments. Each file supports one result or finding in the project.
| File | Model | Purpose | Key result |
|---|---|---|---|
modela-det.pt |
YOLO26s (det) | SPARK-2022 detection | test mAP50-95 / mAP50 = 0.895 / 0.990 |
seg-runA.pt |
YOLO26s-seg | 2026 seg, 2022-pretrained | A/B — identical to B |
seg-runB.pt |
YOLO26s-seg | 2026 seg, COCO-pretrained | A/B — pretraining adds nothing |
comp-fragmented.pt |
YOLO26n-seg (2 cls) | fragmented component labels | mask mAP50-95 ≈ 0.31 |
comp-clean.pt |
YOLO26n-seg (2 cls) | clean component labels | mask mAP50-95 ≈ 0.60 |
classifier.pt |
YOLO26n-cls | 10-class satellite classifier | top-1 0.9973 |
Findings these support
- 2022 pretraining adds nothing when fine-tuning on 60k images (runA ≈ runB).
- Label quality was the biggest lever: fragmented → clean component labels lifted instance mask mAP50-95 from ≈0.31 to ≈0.60.
- Fractional box padding (δ = 0.02) raised mean box IoU from 0.884 to 0.937.
Usage
from ultralytics import YOLO
model = YOLO("modela-det.pt") # or any file above
Data: SPARK 2022 / SPARK 2026 (CVI2, University of Luxembourg).
Citation
@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}
}
@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}
}