pytorch_textboxespp's Introduction

TextBoxes++ with PyTorch

The implementation of TextBoxes++ with PyTorch.

Requirement

pip install --upgrade git+https://github.com/jjjkkkjjj/pytorch_SSD.git
conda install lxml
conda install -c conda-forge shapely

Pre-train

First, download SynthText dataset from official.

Second, convert gt.mat into annotation xml files using synthtext_generator.py.

python synthtext_generator.py {path} -id SynthText

usage: synthtext_generator.py [-h] [-in IMAGE_DIRNAME] [-sm] [-e ENCODING]
                              path

Generate Synthtext's annotation xml file

positional arguments:
  path                  directory path under 'SynthText'(, 'licence.txt')

optional arguments:
  -h, --help            show this help message and exit
  -id IMAGE_DIRNAME, --image_dirname IMAGE_DIRNAME
                        image directory name including 'gt.mat'
  -sm, --skip_missing   Wheter to skip missing image
  -e ENCODING, --encoding ENCODING
                        encoding

Train. See demo/pre-train-SynthText.ipynb.
You can download pre-trained model from here.
Pre-trained model's output example;

Train ICDAR2015

First, download dataset from official.

Second, place annotation .txt and .jpg like this;

├── Annotations (place .txt)
└── Images (place .jpg)

Train. See demo/train-ICDAR2015.ipynb.
You can download pre-trained model from here.
ICDAR's model output example;

Convert png to jpg for Born Digital Images

$ python png2jpg ~/data/text/Born-Digital-Images/Images/ -d
Converting...	100.0%	[307/307]
finished

Reference

pytorch_textboxespp's People

Stargazers

Watchers

pytorch_textboxespp's Issues

loc loss became inf

When ICDARFocusedSceneTextDataset and ICDARBornDigitalTextDataset' was trained, loc loss was inf.
The reason is that "localization height or width = 0"

Recommend Projects