Cadasym is a corpus for Computer Vision on symbols in cadastral maps.
Background: Whenever a Swiss parcel or building changes its geometry, land surveyors are required to submit a so-called “mutation plan” to the local authorities. Today, this is done in a completely digital workflow, but for most of the 20th century, plans were submitted on paper. By analyzing the archived plans, we would like to eventually reconstruct the history how buildings have developed over time. At the moment, the images in the corpus were all taken from cadastral mutation plans supplied by the City of Zürich. In other Swiss municipalities, the plans should look identical, but they will likely not have used the same equipment for scanning paper plans to electronic images.
Purpose: The images from this corpus are useful for testing, evaluating and training computer vision systems. The symbol recognition task appears ideal for training Convolutional Neural Networks with synthetic training data; or maybe it’s enough to go with “old-school” algorithmic computer vision. Whatever solution we end up using, we’ll need to evaluate its quality.
Corpus building: To build the corpus, we wrote an ad-hoc desktop application that extracts image snippet from scanned plans. Human users manually classified the image snippets into one of the categories shown below.
Data download: To download the corpus data, see the ZIP file in Releases.
The released ZIP file contains PNG images, 256×256 pixel in size, where the symbol in question is located at the exact center of the image. Quite often, there are other symbols drawn nearby, or there is an overlapping line. That complication is what makes this an interesting problem. The PNG files are currently in one of these folders:
| Category | Sample |
|---|---|
white_circle |
 |
double_white_circle |
 |
black_dot |
 |
double_black_circle |
 |
small_cross |
 |
large_cross |
 |
triangle |
 |
other |
 |
Note: We’ll likely split the white_circle category into several categories by circle size. Because this is rather trivial for a computer (we can just measure
the circle radius), we’ll do this later. Also, we’ll likely add more categories over time.
Public Domain (CC0-1.0): To the extent possible under law, we have waived all copyright and related or neighboring rights to this work. This work is published from Switzerland.
