1. Basics

1-1. Object detection VS Text detection

• Object detection

  • localization, multiple class
  • Low density, low aspect ratio, general box shapes, size variance $\downarrow$

• Text detection

  • localization, only one class (text)
  • High density, high aspect ratio, unique box shapes, size variance $\uparrow$

1-2. Text box representation

• RECT (Rectangle)
  • $(x_1, y_1, width, height)$ or $(x_1, y_1, x_2, y_2)$

• RBOX (Rotated Box)
  • $(x_1, y_1, width, height, \theta)$ or $(x_1, y_1, x_2, y_2, \theta)$

• QUAD (Quadrilateral)
  • $(x_1, y_1, \dots, x_4, y_4)$

• Polygon
  • $(x_1, y_1, \dots , x_N, y_N)$

2. Taxonomy

2-1. Regression-based VS Segmentation-based

• Regression-Based

  • input image → MODEL → representation
  • [-] Arbitrary-shaped text
  • [-] Extreme aspect ratio
  • TextBoxes ‘18

• Segmentation-Based

  • input image → MODEL → pixel information → process → representation
  • [-] complex post-processing
  • [-] interference between instances
  • PixelLink ‘18

• Hybrid

  • input image → Regression-based MODEL → coarse representation → Segmentation-based MODEL → pixel information → post-processing → output representation
  • MaskTextSpotter ‘18

2-2. Character-based VS Word-based

• Character-Based Method

  • Character region map + Character affinity map → detection result
  • CRAFT ‘19

• Word-Based Method

3. Baseline Model - EAST

• EAST : An Efficient and Accurate Scene Text Detector