Skip to content

FacadeAnnotationProcess

Jump to bottom
John Femiani edited this page Jul 14, 2018 · 8 revisions

Facade Annotation Annotation Process

Overview

This wiki page relates to a continuing effort to increase the size and usefulness of our dataset.

  • 2017 data is labelme.zip (as LabelMe/xml)

  • 2018 data will start with the unannotated images from 2017; but I will be proviing a new set soon that uses overpass to select buildings with useful architectural styles and other tags.

Terminology

Some quick definitions:

vertical

In building-space, this is a direction parallel to the force of gravity. In images it has the usual meaning (y axis).

horizontal

In space, this is perpendicular to the force of gravity. In images it has the usual meaning (x axis)

building-local-coordinates (BLC)

A cartesian reference frame centered at the building, with East(Xblc), North($Yblc), and Up(Zblc) axises. The 'Up' directions would be considered vertical; the east and north directions span a horizontal plane near the building.

face-local-coordinates (FLC)

A cartesian reference frame centered on a face or wall of the building. If you stand with your back against the wall facing away from the building, then the origin of the system is the lowest farthest point to you right. The Xflc axis is in the horizontal plane and also parallel in the plane of the wall, and pointing to your left. The Yflc axis is also in the horizontal plane and pointed away from the wall (in the direction you would be facing). The Zflc axis points Up (the same as Zblc).

Objective of Facade Rectification

  • Rectified facades are easier to annotate; you can draw rectangles around most features instead of polygons

  • Rectified facades are (more) semantically correct; the horizontal lines within the plane of the facade map to horizontal lines of the images, ditto for verticals.

Initial GIS-based rectification

The initial GIS-based rectification was close, but not perfectly aligned. During phase I we determined that annotation was fastest if manually corrected the labels.

Manual Corrections

When shown a new image, for each dominant and viable face, draw a quad:

  1. Draw a horizontal edge near the bottom of the face, starting from the left of the image. Some parts of the face may extend below this edge and that is okay (the ground is not always horizontal). It is important that the edge is horizontal in FLC. The starting and ending points of the line may be occluded; it is OK to guess where they are.

  2. Add a point near the top right of the face. Again some walls have ornamentation or ledges that deviate from the plane of the wall; it is important that within the facade plane, this edge is vertical. You may likely have to guess where to put the point, and it may be out in the sky.

  3. Continue the process to add a horizontal edge as close to the topmost part of the building as you can. The line must be horizontal in FLC.

  4. Verify that the last edge (connecting the first and last point) is the projection of a vertical line.

We consider a face dominant and viable if it is large and clear and mostly-included and mostly-unobstructed in the image.

Note
 I expect this to take about 10-15 seconds per facade image. It can be done on reduced resolution versions of the images if that makes things faster.

Testing Rectification

I plan to provide a tool to let you visualize your results. In the meantime you may just have to learn to visualize.

Continuing the Annotation Process..

Once a sizeable number of facades are outlines, I will re-rectify the images using the quads you have provided.

We will then proceed to annotate features in the following order:

Note
 The order listed is so that I have time to review what we had last time and what can be used to compare against prior art, etc. Let me know before you move on from one feature to another.
Clone this wiki locally