| Images | 1,000 training, 500 test |
|---|---|
| Annotation | Text-line polygons with exactly 14 points (7 along the top edge, 7 along the bottom) |
| Languages | English and Chinese |
| What it tests | Long curved lines, mixed scripts, and whether a method handles line-level rather than word-level output |
| How it is scored | Polygon IoU ≥ 0.5, one-to-one matching (ICDAR 2015 style) |
| Licence | Research use, per the dataset repository |
| Download | https://github.com/Yuliang-Liu/Curve-Text-Detector |
| Paper | Liu et al., Detecting Curve Text in the Wild: New Dataset and New Solution, 2017 (Pattern Recognition 2019) |
SCUT-CTW1500 (Liu et al., 2017) contains 1,500 images with over ten thousand text instances, more than three thousand of them curved. Unlike Total-Text, the unit of annotation is the line, and every polygon has exactly 14 vertices, which made it convenient for methods that regress a fixed number of boundary points.
What it tests
- Long, gently curving lines (banners, circular logos, arched shop names) where the curvature accumulates over many characters.
- Line-level grouping. Word-level detectors must merge words into lines to score well.
- Mixed Chinese and English.
Gotchas
- The 14-point convention encourages methods to emit 14 points; nothing in the metric requires it, but converting a mask to a 14-point polygon can lose IoU.
- Early annotations had some inconsistencies in line grouping that were fixed in a later release; note which version a result used.