| Images | 10,000 training, 10,000 test (MLT 2019); MLT 2017 had 7,200 training, 1,800 validation and 9,000 test images |
|---|---|
| Annotation | Word-level quadrilaterals with script label and transcription |
| Languages | Arabic, Bangla, Chinese, Devanagari, English, French, German, Italian, Japanese, Korean |
| What it tests | Script-agnostic detection, script identification, and recognition across writing systems |
| How it is scored | Precision, recall and H-mean with one-to-one matching at IoU ≥ 0.5 for detection; separate tasks for script identification and end-to-end |
| Licence | Research use; registration on the Robust Reading Competition site |
| Download | https://rrc.cvc.uab.es/?ch=15 |
| Paper | Nayef et al., ICDAR2019 Robust Reading Challenge on Multi-lingual Scene Text Detection and Recognition (RRC-MLT-2019), ICDAR 2019 |
MLT (Multi-Lingual Text) is the Robust Reading Competition’s answer to a simple objection: most benchmarks are English. The 2017 edition covered nine languages; MLT 2019 grew to ten languages across seven scripts and doubled the image count. A synthetic multilingual companion set (SynthText-MLT) is provided for pretraining.
What it tests
- Whether a detector generalises across scripts with very different shapes: connected Arabic, dense CJK, headline-bearing Devanagari.
- Script identification as its own task, since a recognizer must be chosen per script.
Gotchas
- Language imbalance is real; per-script results say more than the overall H-mean.
- The training set is large by scene-text standards, so MLT-trained detectors often transfer well to other benchmarks. Papers that quietly include MLT in training are not comparable with those that do not.