| Images | 63,686 images from MS COCO 2014 |
|---|---|
| Annotation | Axis-aligned word boxes with transcription and attributes (legibility, machine-printed vs handwritten, language); v1 has 173,589 text instances |
| Languages | Mostly English, with other languages tagged |
| What it tests | Scale: text that appears by accident in everyday photos, at every size and quality |
| How it is scored | ICDAR 2015 style IoU at 0.5, computed with the COCO-Text evaluation API; illegible text is don't-care |
| Licence | Annotations under the COCO-Text terms; images under the MS COCO / Flickr terms |
| Download | https://bgshih.github.io/cocotext/ |
| Paper | Veit et al., COCO-Text: Dataset and Benchmark for Text Detection and Recognition in Natural Images, 2016 |
COCO-Text (Veit, Matera, Neumann, Matas and Belongie, 2016) took the MS COCO image set, which was never collected with text in mind, and annotated every piece of text in it. That makes it the closest thing to “text in the wild at internet scale”: most images have no text, some have dozens of tiny words, and a large share is illegible.
What it tests
Recall on incidental text at all scales, and precision in images where nothing should fire. A detector that does well on posed benchmarks can look very different here.
Gotchas
- Boxes are axis-aligned even for rotated text, so polygon detectors are evaluated against loose ground truth.
- Version 2 (2017) re-annotated the set; make sure a result names the version.
- The attribute labels are useful for analysis (how does a model do on handwritten text?) and are often ignored in papers.