Artem Shelamanov A*

home/portfolio/solar filaments

Solar Filaments

Given a 2048 × 2048 image of the Sun in H-alpha light, find every filament and draw its exact outline, one mask per filament.

Kaggle · Filament Segmentation 2026 · in progress

The Sun in H-alpha light, with filaments outlined
expert labels (17) my model (12)
fig. 1 · one training image, scored out of fold: the model that drew these never saw it. 17 filaments labelled, 12 found, 11 matched. it missed the biggest one, bottom centre. picked from the upper middle of the results, not the top.

Filaments are dense clouds of solar material held above the surface by magnetic fields. They sit at the core of solar eruptions, so finding them automatically matters for space-weather forecasting.

The dark threads on the disc are the filaments. Experts outlined them by hand; the job is to do the same automatically, and to get the fine shape right, not just a box around it.

The Sun in H-alpha light, with filaments outlined
fig. 2 · the dashed square from fig. 1 at full resolution. white is the expert, orange is the model.

Image: GONG H-alpha, National Solar Observatory. Expert outlines: the MAGFiLO dataset.

Top 100
on the public leaderboard, October 2026. The competition runs until 15 November.
0.40
panoptic quality, public leaderboard
13
submissions in three days
707
training images, 2048 × 2048 each
01

How it works.

What did not work: classic image processing. Thresholding the dark regions found 510 real filaments and 4,415 things that were not.

  1. 01

    Detect

    A YOLO detector proposes a box around everything that might be a filament.

  2. 02

    Draw

    A small U-Net looks at each box at full resolution and draws the mask inside it.

  3. 03

    Vote

    Five detectors look at the same image and vote on which masks to keep.

02

The climb so far.

Every submission, in order. Not every idea was a good one.

0.320.340.360.380.40Oct 9Oct 10Oct 110.350.40
show as a table
#DayScoreWhat changed
1Oct 90.35First two-stage baseline
2Oct 90.33Stricter thresholds
3Oct 90.34Consensus with a Mask R-CNN
4Oct 90.35Three models, flip augmentation
5Oct 90.35One more variant of the baseline
6Oct 100.37Mask model trained on every annotator
7Oct 100.37Learned quality score
8Oct 100.3977-feature scorer, merged fragments
9Oct 100.39New tile detector
10Oct 110.39Five-detector ensemble
11Oct 110.38Probe: keep more masks
12Oct 110.39Probe: keep fewer masks
13Oct 110.40Rank-averaged scores
Public leaderboard score (Panoptic Quality, higher is better) for each of my 13 submissions.

new best

  1. +0.02 Mask model trained on every annotator Oct 10
  2. +0.02 77-feature scorer, merged fragments Oct 10
  3. +0.01 Rank-averaged scores Oct 11

fell short of the best so far

  1. −0.02 Stricter thresholds Oct 9
  2. −0.01 Consensus with a Mask R-CNN Oct 9
  3. −0.01 Probe: keep more masks Oct 11
03

The code.

The repository goes public once the competition allows it. Until then, the competition page has the task and the data.

Code on GitHub · coming The competition ↗