Checklist USA-NA-AIO 2026 Round 2 · Task 4
Classifying Geometric Shapes from RGB Images
Classify 112×112 images as circle, triangle or rectangle when only 200 of 1,000 training images are labelled.
The task
Problem 4 (40 points, open-ended). Each image contains exactly one shape on a white background, with random position, scale, rotation, geometry and colour. Training images 0–199 are labelled; images 200–999 are unlabelled; 300 test images are unlabelled. Labels: 0 circle, 1 triangle, 2 rectangle.
The design notes describe the task as semi-supervised learning with 20% labelled data, no validation set and pretrained models prohibited.
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 (200 labelled), 300 test images, sample submission.
- You submit
- A prediction CSV, a notebook and a .docx report, each named Shape_Classification_LastName_FirstName_SchoolName.
- Scoring
- Macro F1; invalid or missing labels count as wrong. Baseline = no predictions (score 0). Final score = (X − X_baseline)/(X_best − X_baseline) × 100%.
- Rules
- All pretrained models (torchvision.models, Hugging Face, …) are prohibited, even without their weights; violation scores 0.
- Manual labelling of unlabelled training or test images is prohibited.
- Format
- 2026 USA-NA-AIO Round 2, Day 2 (5 April 2026).