Total-Text

The first widely used curved-text benchmark. Polygons around words on signage, logos and packaging that bend, arc and wave.

last verified 2026-09-05

Images1,255 training, 300 test
AnnotationWord-level polygons with a variable number of vertices, plus orientation labels (horizontal, multi-oriented, curved) and transcriptions
LanguagesEnglish
What it testsCurved and arbitrarily shaped words; whether a detector's output representation can follow the text
How it is scoredPolygon IoU with the ICDAR 2015 one-to-one protocol at 0.5 (most papers); DetEval results also exist
LicenceResearch use, per the dataset repository
Downloadhttps://github.com/cs-chan/Total-Text-Dataset
PaperCh'ng and Chan, Total-Text: A Comprehensive Dataset for Scene Text Detection and Recognition, ICDAR 2017

Total-Text (Ch’ng and Chan, ICDAR 2017) was built to break detectors that assumed text is straight. About half of the images contain curved words, and every word is annotated with a polygon rather than a box or quadrilateral.

What it tests

Whether a detector’s representation can describe the text at all. Box and quad detectors (CTPN, EAST) cannot produce a tight match on a curved word regardless of how well they see it; segmentation detectors (CRAFT, DBNet, FAST) can.

Gotchas

  • Polygon IoU is sensitive to how many vertices a method emits; over-simplified polygons lose IoU on long curves.
  • The dataset is word-level. Line-level detectors need a word-splitting step before evaluation.
  • An updated annotation set (“Total-Text v2”) corrected a number of ground-truth polygons; papers before and after the correction are not exactly comparable.