[2021] Vision Transformer for Fast and Efficient Scene Text Recognition

[osuossu8]

https://arxiv.org/pdf/2105.08582.pdf

[osuossu8]

Train Dataset

use synthetic data

• MJSynth (MJ)
  • 8.9M
  • 1,400 different fonts
• SynthText (ST)
  • 5.5M
• In the STR framework, each dataset contributes 50% to the total train dataset. Combining 100% of both datasets resulted to performance deterioration

[osuossu8]

Setting

• 224 × 224 image for DeiT
• 2080Ti GPU
• Data Augmentation : RandAugment