From: First-person reading activity recognition by deep learning with synthetically generated images

Process | Parameters | Values |
---|---|---|

– | Canvas of each page | \(210\sqrt {2} \times 210\) pixel by appropriate scaling |

T
| Blank space | 10% at the top, bottom, left, and right in the canvases |

Columns | 1 or 2 | |

Pages containing figures | 80% of the entire generated pages | |

Place to put figures | The top or the bottom in a column | |

Category of figures | Mathematical figures, tables, and general pictures | |

Figure size | Height: 12.5∼50%, or 100% of the column height
| |

Width: fixed at the same length of the column | ||

Size of headlines | 10% of the page height | |

Place to put headlines | Somewhere outside the figure areas | |

Place to put texts | Entire areas in the column except the figures and headlines | |

Text format | Japanese characters in random order (without any rules) | |

Line-breaking | Done with prob. 1% every time putting a character | |

D
| Distortion strength |
α∈[0.1,0.3] |

P
| Homography matrix | Determined as shown in Fig. 4 |

R
| Rotation angles | In [ − 10°,10°] |

– | Resize | The long side shrinks to 256 pixel with keeping the ratio |