Discord

Checklist EUROAI (CEOAI) 2026 Practice Round 1 · Task 3

Star Observatory

Locate a star in a 128×128 image and predict its calibrated flux from the image and atmospheric metadata.

  • Vision
  • Tabular
  • Keypoint localisation and regression

The task

Ground-based observations are degraded by airmass and atmospheric turbulence (Fried parameter r0). Each 128×128 image contains exactly one star; training metadata give image_id, fried_parameter, airmass and target_flux, while the test metadata contain only image_id.

Subtask 1 (20 pts) predicts the star centre; Subtask 2 (80 pts) predicts the true (exponential-scale) flux.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
1,000 training images with train.csv; 300 test images with test.csv (image_id only).
You submit
CSV with 600 rows (subtaskID, datapointID = image filename, answer): '(x, y)' for subtask 1, a float for subtask 2.
Scoring
Subtask 1: MAE (Manhattan distance); 20 pts if < 5, linear 10→0 between 5 and 10, 0 above 10. Subtask 2: RMSE; 80 pts if ≤ 280, linear 80→40 up to 400, 40→0 up to 880, 0 at ≥ 880.
Rules
  • At most 45 minutes notebook runtime (training + inference) on a free 16 GB GPU (T4 on Colab or P100 on Kaggle).
  • No use of LLMs for code writing or submission generation.
  • No pretrained models.
  • No data other than the provided dataset.
Format
Online Practice Round 1, 15 March 2026, 13:00–19:00 UTC (6 hours), open to everyone on Nitro AI Judge; tasks proposed by volunteers and reviewed by the International Scientific Committee.

Details

Year
2026, Online
Round
Practice Round 1 · Task 3
Language
English
License
Not stated by the source