Checklist EUROAI (CEOAI) 2026 Contest Day 2 · Task 3
Odd One Out
Among three AI-generated images in different art styles, find the one that does not contain the object type shared by the other two.
The task
Each datapoint is a set of three AI-generated scenes, each containing three everyday object types from a fixed vocabulary and rendered in a different art style (anime, watercolour, oil painting, cyberpunk, low-poly 3D, pixel art, ...). Exactly one object type appears in two of the images; the odd image is the one without it. Whole-image similarity is deliberately insufficient.
Only three bundled pretrained encoders may be used (CLIP ViT-B/32, DINOv2-base, ImageNet ResNet-50), zero-shot, as frozen feature extractors or fine-tuned on the provided training set.
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
train.zip(train/images/train_XXX_{0,1,2}.jpg,labels.csv,meta.jsonlwith target object, styles and objects per image);starter_kit.zip(predict.ipynband the bundled models).- You submit
- A cloudpickled Predictor object run on Nitro Contestant Cloud; it receives pickled JPEG triples and returns a NumPy int array of odd-image indices (0, 1 or 2).
- Scoring
- Weighted F1 over the predicted index. Points: 5 if F1 < 0.359; 5 + floor(95 × (F1 − 0.359) / (0.90 − 0.359)) for 0.359 ≤ F1 < 0.90; 100 if F1 ≥ 0.90. A crashing submission scores 0 and does not count against the limit.
- Rules
- Only the three bundled pretrained models; no other pretrained models (no other VLMs, detectors, segmenters or backbones).
- No internet in the judging sandbox; no external data or answer keys.
- Submission limit 30, two final submissions (platform settings).
- Format
- Contest Day 2, 17 July 2026 (Nitro AI Judge 07:10–12:50 UTC), on-site in Cluj-Napoca; three tasks, 100 points each.