Detection and recognition of texts in cartographic images

Historical maps give an idea of growth and development of an area. These maps are now available in digitized form. But none are catalogued and annotated. Thus, the performance of any algorithm running on these maps couldn't be evaluated. As a first phase to this project, we have annotated a publicly available map image database.

Annotated images

In the following panel you can see the original images from the David Rumsey Collection and the corresponding annotated ones (the files are lined up side by side).

map1 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1 map2 map1

Data

  • The complete list of images can be obtained in images
  • The complete list of word annotations can be obtained in annotations
  • The complete list of approximate character annotations can be obtained in approximate character annotations
  • Codes

  • The code for the cascaded faster RCNN (detection model) can be obtained from C-FRCNN
  • The code for the PHOC Net (recognition model) can be found in PHOCN
  • Results of different methods on each maps

    Please go to the following link to access all the results. Results

    Details about the data

    We gratefully acknowledge the David Rumsey Map Collection as the source of the map images. It comes with the following notice:

    Images copyright © 2000 by Cartography Associates. Images may be reproduced or transmitted, but not for commercial use. ... This work is licensed under a Creative Commons [Attribution-NonCommercial-ShareAlike 3.0 Unported] license. By downloading any images from this site, you agree to the terms of that license.

    Data Format:

    Each of the annotations are provided as a numpy file. The file contains a python dictionary of python dictionaries. To import the data in python use the command:

    A = numpy.load('filename').item()

    Each dictionary contains the following three fields:
  • vertices: Contains a list of (x,y) values of the position of the vertices according to the image co-ordinate system.
  • name: The label for the polygon. Typically this is word which is bounded by the polygon. In ineligible cases this is 'unreadable'.
  • link_to: This field, if present, is used to link characrters or words. In cartographic images, scenarios like spaced out words or broken words can be found. This field allows us to capture such discontinuity.
  • File to visualize the data against each image:

    Visualizer