Discord

Checklist AICC 2026 Round 4 · Task 1

Sticky Note Blindness

Recover the true object class from images under a typographic attack (a sticky note with a misleading word), using only CLIP ViT-B/16.

  • Vision
  • Multimodal
  • Robust zero-shot image classification (defence against typographic attacks)
Read the official statement

The task

An object-classification system based on CLIP (ViT-B/16) predicts the class whose text embedding is closest to the image embedding. It works on clean images but is vulnerable to a typographic attack: a sticky note bearing a misleading word is attached to the object, and CLIP often predicts the written word.

The images are attacked RGB images at original resolution. The provided class embeddings are L2-normalised CLIP text embeddings computed from class names only, for about 800 labels comprising true object classes and attack words. CLIP achieves 100% zero-shot accuracy on the corresponding clean images, so every misclassification is caused by the attack. Metadata give the fraction of each image covered by the sticky note.

For each attacked image the contestant predicts the integer index of the true object class.

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
data/images/images/ (attacked PNG images), class_embeddings.pt ([K, 512] float32 CLIP text embeddings) and metadata.csv (sample_id, postit_area_pct).
You submit
submission.csv with columns sample_id (0 … N−1) and label (integer class index); other formats score 0.
Scoring
Classification accuracy. Baseline about 0.46, reference solution about 0.90 or higher.
Rules
  • openai/clip-vit-base-patch16 is the only model allowed; no other pretrained models or external data.
  • Individual participation (maximum team size 1); at most 30 submissions per day.
  • AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
Format
AICC Round 4, online on Kaggle, 20 Feb 2026 18:00 UTC – 22 Feb 2026 18:00 UTC.

Notebook

Competition

Solutions

Details

Year
2026, Online (Kaggle)
Round
Round 4 · Task 1
Language
English
License
Varies by task: Sticky Note Blindness — CC BY-NC-ND 4.0; Alchemy and Extreme Condensation — 'Subject to Competition Rules' (Kaggle licence field). Solutions repository: MIT., as stated by the source