| Images | 1,000 training, 500 test |
|---|---|
| Annotation | Word-level quadrilaterals (four points) with transcriptions; unreadable words marked ### and ignored |
| Languages | English |
| What it tests | Multi-oriented, low-resolution, motion-blurred text at natural scale |
| How it is scored | Precision, recall and H-mean (F) with one-to-one matching at IoU ≥ 0.5; ### regions are don't-care |
| Licence | Research use; registration on the Robust Reading Competition site |
| Download | https://rrc.cvc.uab.es/?ch=4 |
| Paper | Karatzas et al., ICDAR 2015 competition on Robust Reading, ICDAR 2015 |
ICDAR 2015 (Challenge 4, “Incidental Scene Text”) was captured with Google Glass while people walked around, without pointing the camera at anything in particular. Words are small, often rotated, and frequently blurred. It replaced ICDAR 2013 as the headline benchmark because it was hard: CTPN scored 0.61 F here against 0.88 on the focused set.
What it tests
- Orientation: the quadrilateral annotations punish axis-aligned detectors.
- Scale: many words are under 20 px tall.
- Robustness to blur and perspective.
Gotchas
- Evaluation ignores “###” regions, but a detector that fires on them is not penalised either; precision numbers therefore depend on how a paper handles don’t-care matches.
- Most published results use a fixed test resolution (often the long side resized to 1,280 or 2,240 px) and single-scale testing. Multi-scale testing adds one to three points and is not always disclosed.
- Test-set labels are available, so treat any “fine-tuned on ICDAR 2015” number with the usual caution about hyper-parameter selection on the test split.