https://github.com/kentaroy47/textsnake_pytorch

Unofficial implementation of textsnake. Mostly to practice code reading.
https://github.com/kentaroy47/textsnake_pytorch

ocr pytorch text-detection

Last synced: 3 months ago
JSON representation

Unofficial implementation of textsnake. Mostly to practice code reading.

• Host: GitHub
• URL: https://github.com/kentaroy47/textsnake_pytorch
• Owner: kentaroy47
• License: mit
• Created: 2020-04-28T02:57:35.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2020-05-14T01:34:41.000Z (over 5 years ago)
• Last Synced: 2025-03-22T18:46:21.424Z (7 months ago)
• Topics: ocr, pytorch, text-detection
• Language: Jupyter Notebook
• Homepage:
• Size: 2.73 MB
• Stars: 2
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# textsnake_pytorch

Unofficial implementation of textsnake.

Most of the codes are borrowed from [TextSnake.pytorch by princewang1994](https://github.com/princewang1994/TextSnake.pytorch), which is a great repo.

## Some new features

* Added resnet50 backbone.

* Added Batchnorm-upsampling blocks for faster convergence.

* Synthtext dataset conversion script.

# What is this repo?

This repo breaks down the 1) dataset setup, 2) model setup, 3) training setup, 4) evaluation/Inference **mostly for code reading purposes.**

I added some comments in Japanese to get some understanding of how TextSnake works.

`1. Prepare Dataset.ipynb` shows how the annotation data are converted to the TextSnake format step-by-step.

`2. Prepare Model.ipynb` shows how to setup the TextNet model.

`3. Train Model.ipynb` shows how to train the model using the prepared dataset.

The model is pretrained on SynthText and finetuned on totaltext.

`4. Inference and Evaluate.ipynb` shows how to visualize and evaluate the TextNet results.

## Results

resnet50 

`Precision = 0.7379 - Recall = 0.6681 - Fscore = 0.7012`

vgg16