EAST Detects Text in One Shot
EAST predicts word boxes and their angles straight from pixels, so tilted and long lines come out in one forward pass.
Why Does This Exist?
Shop signs tilt and receipts run long. Axis-aligned detectors either clip letters or swallow neighbours, and multi-stage text pipelines were slow.
EAST predicts dense score and geometry maps in one pass. It specializes text detection, borrows box logic from object detection, and feeds clean crops to optical character recognition.
Think of It Like This
Tracing a tilted label with a ruler
You lay tracing paper over a tilted jar label, shade where ink sits, then draw one rotated rectangle around the shading with a protractor angle.
It fails on curved bottles. One straight rotated edge cannot hug letters that bend around the glass.
How It Actually Works
For each pixel, predict score for text presence, where near 1 means text. Where is high, regress distances to four box edges plus angle , where is box rotation in radians. A pixel with and votes a box tilted about 30 degrees, since .
Worked example
A tilted word returns above threshold 0.8, with edge distances and . The box corners rotate by 0.52 radians and enclose all letters. A straight box on the same pixels cuts the last letter and scores IoU 0.61 versus 0.89 for the rotated fit. The 0.28 gap is pure angle.
Watch Out For
Merging nearby lines into one quadrilateral
Close receipt lines join on low-resolution maps. The symptom is one tall box covering two lines. Raise the score threshold and shrink the geometry field near line gaps instead of lowering NMS alone.
The Quick Version
- EAST outputs a score map plus rotated geometry per pixel in one pass.
- Fixture angle 0.52 radians tilts the box about 30 degrees to fit the word.
- Rotated IoU 0.89 beats straight 0.61 by 0.28 on tilted text.
- Long and tilted lines work well, curved text does not.
- Tune score threshold to stop line merging before touching NMS.