Pixel-Based Ground Truth (GW20) — Sophie Li

From CSclasswiki
Jump to: navigation, search


Fall 2017

fixErrors.m: Use to correct errors in generated page ground truth.

  • Parameters
    • Page ground truth
    • Optional: Grayscale page images (same dimensions as page ground truth)
  • Movement Controls
    • Left click/right click: Zoom in/zoom out
    • Arrow keys: Pan current screen
    • Comma (back) and period (forward) keys: Scroll through pages (cells in cell array)
    • H: Shows help in console
  • Display Settings
    • Semicolon key: Toggle black and white colormap
    • Quote key ('): Toggle jet colormap
    • A: Zoom out and display all characters
    • Left and right brackets: Scroll through all character codes in the current page
    • C: Displays given character codes (dialog input)
    • Q: Displays character code on next mouse click (useful for remembering character codes)
    • B: Toggles grayscale original page image
      • Only available when using the optional grayscale page images as the second argument
      • Scrollbar at bottom of window adjusts the display/binarization threshold
  • Editing Controls
    • F: Sets fill character code (default is 99)
    • L: Calls freehand lasso tool to select/fill areas
    • P: Calls polygonal lasso tool to select/fill areas
    • M: Sets mode type (0, 1, or 2*, default is 0)
      • 0: Fills selection and preserves current letter boundaries
      • 1: Fills selection, regardless of current letter boundaries
      • 2: Fills selection, and preserves the letter boundaries respective to the grayscale page image and its current threshold
        • Only available when using the optional grayscale page images as the second argument
        • Used when ground truth binarization deviates from the actual image
  • Saving/Quitting
    • Escape: Exits function, and returns ground truth without edits
    • Space bar: Exits function, and returns ground truth with edits


GW20CharTruthJanuary2018.mat: Edited page ground truth (From GW20CharTruthJune2017.mat), with the addition of 61st character (ampersand).

Spring 2018

02/02 — 02/08

  • Fixed bugs in fixErrors.m related to lag and character cycling
    • hold on preserves old plots, use cla before showing updated plot to prevent lag

02/09 — 02/15

Papers:

Page Segmentation of Historical Document Images with Convolutional Autoencoders

PCA-Initialized Deep Neural Networks Applied To Document Image Analysis

Historical Document Image Segmentation with LDA-Initialized Deep Neural Networks

A Pitfall of Unsupervised Pre-Training

02/16 — 02/22

DIVA-HisDB: A Precisely Annotated Large Dataset of Challenging Medieval Manuscripts

Data-Dependent Initializations of Convolutional Neural Networks