What this is
Drop an image into the detector and a modern model boxes every line of text and reads it, entirely in your browser. Around it sit the model pages, which put a decade of scene text detectors side by side with the numbers each paper reports, the dataset pages, a benchmarks explainer, and guides with code you can run. Every number names its source; values quoted from tables stay marked as unverified until someone checks the PDF.
Who runs it
textdet.com is an independent project. It is not affiliated with PaddleOCR, whose models the detector uses, or with the authors of the papers documented here; they are credited wherever their work appears, and every project the site runs on is listed on the credits page. Contact details will be published here shortly.
How it is built
Static pages, one small edge worker, and a detector that runs through ONNX Runtime Web with WebGPU where available and WebAssembly elsewhere. The detection and recognition models are PaddleOCR's Apache-2.0 releases packaged by the MIT-licensed ppu-paddle-ocr SDK; the fonts are Bricolage Grotesque and IBM Plex, self-hosted. No image, text or result you produce in the detector is ever sent anywhere; see privacy.
Corrections
If a number, a licence or a description on any page is wrong, or if you are an author who would like something about your work changed, tell us and it will be fixed. The changelog records what changed and when.