textdetDetect text
Datasets2016

SynthText

Around 800,000 synthetic images with about 8 million rendered words, generated by pasting text into natural scenes with geometry-aware placement. The pretraining set almost every detector starts from.

last verified 2026-09-05

ImagesAbout 800,000 synthetic images
AnnotationWord- and character-level boxes with transcriptions, produced by the renderer
LanguagesEnglish
What it testsNothing: it is training data, not a benchmark
How it is scoredNot used for evaluation
LicenceResearch use; see the generator repository
Downloadhttps://github.com/ankush-me/SynthText
PaperGupta, Vedaldi and Zisserman, Synthetic Data for Text Localisation in Natural Images, CVPR 2016

SynthText (Gupta, Vedaldi and Zisserman, CVPR 2016) solved the labelled-data problem for text detection by manufacturing it. The generator estimates depth and surface regions in a background photograph, chooses fonts, colours and effects, and renders text onto plausible surfaces with perspective, producing exact word and character boxes for free.

Why it matters

  • Almost every detector since 2016, including CRAFT and DBNet, pretrains on SynthText before fine-tuning on a few hundred or thousand real images.
  • It is the only large source of character-level boxes for Latin text, which is what made CRAFT’s weakly supervised training possible.

Gotchas

  • The archive is about 41 GB.
  • Synthetic fonts and placements have a distribution of their own; models pretrained here and never fine-tuned do badly on curved and handwritten text.
  • Later generators (SynthText3D, UnrealText, and diffusion-based pipelines) produce more realistic data; results that name only “synthetic pretraining” may not mean this set.