Run it
Drop / choose your own image
–regions · angles below
backend –latency –
← All models · Scene text detection
Wild · object-detection (text) · WASM
Real signage isn't neatly horizontal. Because the decode fits a minimum-area rectangle to each region, the boxes tilt to follow slanted poster type — and each one reports its recovered angle. Crank the resolution up for dense text.
–regions · angles below
The model outputs a probability map that already follows the ink, whatever its angle. The decode groups the bright pixels into a region, computes its convex hull, then finds the tightest rectangle over any rotation (rotating calipers). Axis-aligned boxes would swell to fit a 45° headline; an oriented box hugs it — and hands you the angle for free, which is what a deskew-then-OCR pipeline needs.