Scene Text Recognition

[ptpuyen1511]

TL;DR

Some notes from VINAI's seminar: Scene Text Recognition

Author: Minh-Hoai Nguyen

Key Takeaways

Dataset

• SynthText:
  • 80k synthetic images from 8k background images
• ICDAR15
  • 720x1280 images captured by Gooogle Glasses
  • 1k train, 0.5k tests images
• ICDAR17
  • 9 languages, 6 scripts
  • 7.2k train, 1.8k validation, 9k test images
• Coco-Text
  • 63k images with 173k text annotations
• Total-Text
  • Text of various shapes: horizontal, multi-oriented, curved
  • 1255 train, and 300 test images
• CTW1500
  • curved text
  • 1k train, 0.5k test
• Chinese Street View
  • 30k fully annotated
  • 400k weakly annotated
• MSRA-T500
  • 300 train, 200 test English and Chinese
  • Text instances are labeled at text-line level

Three recent papers

Real-time Scene Text Detection with Differentiable Binarization (AAAI 2020)

Convolutional Character Networks, ICCV'19

Colab: https://www.dropbox.com/s/mxvoku6rh8bzq06/CharNet.zip?dl=0

ABCNet: Real-time Scene Text Sptting with adaptive Bezier-Curve Network