Checklist ROAI 2026 Selection Camp – GPU Practical Round · Task 1
What do you see
Generate five short natural-language descriptions for each test image.
The task
The dataset pairs images with several short, semantically similar alternative captions. For every image in the test set the contestant must generate exactly five textual descriptions of its visual content.
Generated captions are compared with the reference captions using BLEU-1 and BLEU-2.
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
- Images paired with multiple short reference captions (training); test images without captions.
- You submit
- For each test image, a list of exactly five caption strings, submitted in the judge's CSV format.
- Scoring
- M = (BLEU-1 + BLEU-2) / 2. M < 0.11 gives 5 points; 0.11–0.62 is interpolated linearly from 5 to 60; 0.62–0.665 from 60 to 75; 0.665–0.685 from 75 to 90; 0.685–0.710 from 90 to 95; M > 0.710 gives 100.
- Rules
- GPU round
- On-site at the National College of Informatics "Tudor Vianu", Bucharest
- Internet access and packages restricted to the allowed list in the round rules
- Final score computed on hidden test data; only selected submissions count
- Format
- ROAI 2026 selection camp, GPU practical round, 27 May 2026 (06:00–12:00 UTC on the judge).