| Images | 229 training, 233 test |
|---|---|
| Annotation | Word-level axis-aligned rectangles with transcriptions |
| Languages | English |
| What it tests | Horizontal text at readable scale; the easy case every detector should saturate |
| How it is scored | DetEval (one-to-one, one-to-many and many-to-one matching) in the original competition; many later papers report the ICDAR 2015 IoU protocol instead, so check which one a table uses |
| Licence | Research use; registration on the Robust Reading Competition site |
| Download | https://rrc.cvc.uab.es/?ch=2 |
| Paper | Karatzas et al., ICDAR 2013 Robust Reading Competition, ICDAR 2013 |
ICDAR 2013 (Challenge 2, “Focused Scene Text”) is the benchmark most 2016-era detectors were tuned on. Photographs were taken with the text in mind, so words are large, roughly horizontal and in focus.
What it tests
Baseline competence: can a detector find clean horizontal words? A modern model should score in the high 0.8s or better. Numbers near 0.88 F (CTPN, 2016) were state of the art then and are unremarkable now.
Gotchas
- Two scoring protocols circulate. The competition used DetEval, which rewards one-to-many matches; papers that report the IoU-based ICDAR 2015 protocol on this dataset get lower numbers for the same boxes. A table that mixes both is comparing apples with pears.
- The training split is tiny (229 images). Everyone pretrains on SynthText or a larger real set first.
- The images are a subset of ICDAR 2011 with corrected annotations; do not train on 2011 and test on 2013.