Discord

Checklist ONIA 2026 Team Selection Camp (Extended Squad) – Test 1 · Task 1

Glitch Hunter 2: G3P1T Reloaded

Romanian title: Vânătorul de Glitch-uri 2: G3P1T Reloaded

Segment four classes of synthetic glitches in corrupted images and count the glitch instances of each class, while ignoring decoy artefacts.

  • Vision
  • Semantic segmentation and instance counting
  • Romanian original
Read the official statement (Romanian)

The task

Corrupted 512×512 images contain glitches of four types: noise (random RGB pixels), invert (inverted colours), solid (single-colour blocks) and scramble (pixels shuffled inside a rectangle, leaving the colour histogram unchanged). Decoys that must be left unmarked are also present: low-quality JPEG re-encoded regions, hue-shifted areas and Gaussian-blurred patches.

For each test image the contestant submits a single-channel class mask (0 = clean, 1–4 = glitch class) and, in instance_counts.json, the number of distinct glitches of each class.

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

At a glance

You get
chaos_returns_dataset.zip: train/ with 200 corrupted images, 200 class masks (values 0–4) and instance_counts.json; test/ with 200 corrupted images. Archive of 79,822,779 bytes.
You submit
submission.zip containing masks/img_XXX.png (200 uint8 PNGs, same size as the inputs) and instance_counts.json at the root.
Scoring
Final = 0.7 × mean macro-IoU over classes 1–4 (IoU = 1 when a class is absent in both prediction and ground truth) + 0.3 × mean count score, where count score = 1 / (1 + |predicted − true|); scaled to 0–100. The partial score shows segmentation only.
Rules
  • On-site, Timișoara (West University of Timișoara and Politehnica University of Timișoara)
  • No internet access; only preinstalled offline resources (NLTK, spaCy en_core_web_sm, scikit-learn, Gensim glove-wiki-gigaword-50, fastText without pretrained models, and the Hugging Face models distilbert-base-uncased, MoritzLaurer DeBERTa-v3-large-mnli-fever-anli-ling-wanli, roberta-large-mnli, all-MiniLM-L6-v2 and whisper-small)
  • A GPU Jupyter environment hosted by UVT was provided during the tests
  • Problem statements unlocked with a password handed out in the room
Format
ONIA 2026 Lot Lărgit, test 1, 23 May 2026 (06:15–11:20 UTC on the platform).

Statement

Competition

Details

Year
2026, Timișoara, Romania (UVT + UPT)
Round
Team Selection Camp (Extended Squad) – Test 1 · Task 1
Language
Romanian
License
Not stated by the source