Checklist AICC 2026 Round 8 · Task 1
Dice Counting
Count the dice and the total of up-facing pips in photographs whose dice counts lie beyond the training range.
The task
A board-game club photographs the dice left on the table and records how many dice are present and the total of the up-facing pips. The ledger is blank for the 50 newest photos, and the task is to fill it in.
There are 250 RGB JPEG photos of six-sided dice, each 2736 × 2736 pixels, with dice of varied material, finish, size and colour, from well-separated sets to crowded piles. The 200 training photos contain 1 to 10 dice (pip totals 1 to 37) and are labelled; the 50 test photos contain 10 to 25 dice (pip totals 22 to 96). The test set is out of distribution by design, and only one training photo is as crowded as the test photos.
Subtask 1 (30 points) asks for the number of dice in each test photo; subtask 2 (70 points) asks for the total number of up-facing pips.
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/ and test/ photos (in the starter kit, with
baseline.ipynb),train_labels.csv(image_id,n_dice,total_sum) andtest.csv. - You submit
submission.csvwith columns subtaskID, datapointID, answer: 100 rows,count_{image_id} rows for subtask 1 andsum_{image_id} rows for subtask 2, each with a whole-number answer.- Scoring
- For each subtask, the per-photo absolute error is capped (12 for the count, 36 for the sum; missing or non-numeric answers count as the cap) and averaged over the 50 photos (MAE). Count points: 0 for MAE >= 4.0, linear 0→12 from 4.0 to 2.0, linear 12→30 from 2.0 to 0.25, 30 for MAE <= 0.25. Sum points: 0 for MAE >= 16, linear 0→20 from 16 to 6.5, linear 20→70 from 6.5 to 1.2, 70 for MAE <= 1.2. Points are rounded to whole numbers; maximum 100.
- Rules
- Pretrained weights are not allowed.
- External datasets are forbidden; no additional dice photos may be added.
- Each submission must include the source code as a .py file (maximum 35 KB) and the predictions CSV (maximum 50 MB); organiser-provided seed-setting code must be used.
- Round rules (honour system): no LLMs or AI assistants for code or ideas; no internet beyond official library documentation; no communication or sharing of approaches during the contest; individual participation only.
- Format
- AICC Round 8 (special edition following the CEOAI syllabus), online on Nitro AI Judge; open from 12 Jun 2026 17:00 to 14 Jun 2026 17:00 (time zone not stated on the platform); contestants are asked to work within a single self-chosen 6-hour session.