TextField

The code behind a TIP 2019 paper on detecting irregular scene text, using a deep direction field, with Caffe prerequisites, two datasets, and trained models with reported scores.

TextField is the implementation of the paper Learning A Deep Direction Field for Irregular Scene Text Detection, published in TIP 2019, with an arXiv link in the README. The method targets irregular scene text, the cases where rectangular or quadrilateral detection approaches struggle, and uses a direction field representation to handle them. The repository is presented as the code and trained models accompanying the published paper.

The prerequisites are Caffe and a SynthText pretrained model, with a download link, plus two datasets: Total-Text and ICDAR2015, each linked. Total-Text covers irregular text shapes, which lines up with the method's focus, and ICDAR2015 is a standard benchmark for scene text detection. Both are common fixtures in this corner of research.

A results table lists recall, precision, and F-measure with Google Drive links to the trained models. The first row reports 0.816 recall, 0.824 precision, and 0.820 F-measure, and the second reports 0.811 recall, 0.846 precision, and 0.828 F-measure. The numbers sit close together, which is typical for the two runs shown.

The README stays lean: paper reference, prerequisites, datasets, and trained models, with no training or inference commands included. Reproduction rests on the documented pieces, and the arXiv page carries the method description in full. Anyone picking this up should expect a research codebase oriented around the paper rather than a turnkey tool.

This is a research repo in the classic style: paper reference, prerequisites, datasets, and trained models, with the results table doing the talking. The arXiv link carries the full method for anyone who wants the details.

• Source 1 · Project repository
• Source 2 · README