Checklist IOAI 2025 Individual Contest · Task 4
Restroom Icon Matching
Train a matching model that, for each cropped restroom icon, finds the opposite-gender icon from the same restroom in a gallery.
The task
Restrooms usually show a pair of male and female icons drawn in a shared style. The task is to learn that link: given a query icon, find its cross-gender partner from the same restroom, using a model you train.
The training data covers 82 restrooms, each with four PNGs: cropped and original versions of the male and female icon, all named by restroom ID. In the hidden validation and test sets, cropped query icons are matched against a gallery of original icons, with filenames renumbered and shuffled so IDs give nothing away.
Every query has exactly two gallery images from its restroom (one male, one female), and the correct answer is the one of the opposite gender. The validation set has 10 queries and the test set 30. A CLIP model was provided and is used by the baseline.
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
- Training set of 82 restrooms × 4 PNG icons (crop/orig × female/male). The validation (
test_a) and test (test_b) sets each have query/ and gallery/ folders, with 10 and 30 queries and twice as many gallery images. - You submit
submission_a.npy(validation) andsubmission_b.npy(test): 1-D arrays holding one gallery image ID per query, produced bysubmission.ipynb, which must include model training and matching.- Scoring
- Score = number of correct matches / total queries (0 to 1); submissions that exceed the time limit score 0. The Scientific Committee's highest score (0.90) and baseline score (0.77) on Leaderboard B were used for score normalisation.
- Rules
- Notebook must finish within 15 minutes on the test machine; over-time submissions score 0
- At most 15 successful submissions; up to 2 can be selected and the higher Leaderboard B score counts
- Only a notebook may be submitted (no self-made mounted datasets or .pth files)
- Test machine has no internet access; a CLIP model is provided, and other pretrained models may be used only if they import offline
- Format
- Individual Contest, Day 2 (6 August 2025); six hours for the day's three tasks