SPARK-2026 — 20-class spacecraft component segmentation

A single compact YOLO26n-seg model that classifies, detects and fine-grained segments spacecraft components in one pass.

Approach

The multi-task problem is recast as a single 20-class instance segmentation task: each satellite is decomposed into its components — 10 satellites × {body, panel}.

• classification is read off the predicted component classes (no separate classifier)
• detection boxes are the per-satellite union of component boxes (padded by δ = 0.02)
• segmentation is the body / panel masks directly

Results

S_acc = 0.20·S_cls + 0.40·S_det + 0.40·S_seg · metrics measured on a held-out 2k-image validation subset.

Classes (20)

satellite × {body, panel}, indices 0–19 — satellite order: Cheops, LisaPathfinder, ObservationSat1, Proba2, Proba3, Proba3ocs, Smart1, Soho, VenusExpress, XMM Newton (index = sat_idx*2 + comp, body=even, panel=odd).

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

path = hf_hub_download("huseyincavus/spark-2026-comp20-seg", "best.pt")
model = YOLO(path)
results = model("satellite.jpg", imgsz=640, conf=0.25, retina_masks=True)

Training

YOLO26n-seg · 46 epochs · 80k images (60k train / 20k val) · 640px · 2×T4.

Data

SPARK 2026 (Stream 1) spacecraft imagery; the 20-class labels were derived from the released component masks (red = body, blue = panel). Datasets © CVI2, University of Luxembourg.

Limitations

• Component panel masks are harder (thin structures) — 0.752 vs 0.910 for bodies.
• Detection is reported as an IoU-thresholded hit-rate proxy, not a full PR-curve AP.
• Long solar arrays are predicted as multiple panel segments (matches the training labels).

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}
}