Some Frames From A Prototype Segmentation Interface

This page shows a few frames captured from a prototype graphical interface for hand editing segmentations of aerial images.
The intent is to give you the basic idea of how you would interact with the tool. The page is organized as a sequence of images that show some of the variations of initial segmentations to edit, followed by several editing steps on one of the initial segmentations:

Original image to be segmented
3 different initial segmentations generated by a computer program
3 steps in editing one segmentation

Note that the process of capturing and redisplaying the images to make this page has significantly degraded the quality of the images shown here compared to what the the user of the tool sees. In particular, there is a great deal more detail visible in the original grey scale images.

Each image from the interface shows the original grey scale image with segment boundaries laid over it. There are 3 colors of lines:

Red lines show the segmentation originally drawn by a biologist hand segmenting original prints of aerial photos with mylar overlays. These lines were later digitized using a mouse and saved as a GIS layer.
Blue lines show an overly detailed segmentation automatically created by a computer program.
Green lines show pixels where the two segmentations are in perfect agreement.

In the real interface, only the blue lines would be visible because the user would be trying to generate a segmentation rather than match one. The other lines are only shown here to give an indication of what we are trying to generate.

Initial image to segment

This is a small piece of a 1:24,000 black and white aerial photo taken in Lake Clark National Park, Alaska in 1993 as part of a study relating to salt marshes. It is the base photo used throughout the example on this page. Again, remember that there is a lot of detail lost in preparing the image for use on this web page.

Note that not all of this image was of interest to the biologist who originally segmented it. In the segmentations that follow, the parts of the image that were not segmented by the biologist are blacked out.

Primitive segmentation

Here is an example of a segmentation that could be used as an initial input to the hand segmentation aid. It was created using a segmentation algorithm that has a tendency to grossly oversegment an image. While it has far more segments than we would like for the final segmentation, it has several nice properties for use in this tool:

The segmentation is easy to compute.
All segments have fully closed boundaries so no boundary completion algorithms are required.
Most boundaries that we would like to see in the final segmentation appear somewhere in this segmentation.
We just need to remove the excess ones.
We can apply various algorithms to selectively make the segmentation coarser
(see following segmentations).

A higher level initial segmentation

This is an example of the results of applying a reconstruction algorithm to the primitive segmentation so that we can start from a segmentation with fewer regions if we want. This is the segmentation that we will start from in the examples later in the page.

A much higher level initial segmentation

This is another reconstruction example to show what happens as we keep raising the level of coarseness. We won't use this in the examples below because it is too coarse for what we want, but we could use it (or any other fully closed segmentation) if that was appropriate to the our application.

First stroke with user interface

We now examine a few snapshots from applying the interface to the the 2nd segmentation shown above. To minimize time spent waiting for web downloading, I have not shown the original segmentation next to the interface, but you can open another browser on this page if you find it useful to see them side by side.

In the segmentation interface, the user holds the mouse button down and drags the mouse across any segment boundary that should be deleted. As the mouse crosses each boundary, the segment "pops" like a balloon. This image shows the result of dragging the mouse in a single stroke along the axis of the long grey region in the upper right of the image.

Continuing cleanup of same region...

A few more strokes of the mouse in the same region yields the nearly complete segmentation of the full region. The small spines that you see left along the edges of the region are artifacts of the prototype. Since it is still in its early stages, these spines are not removed until the segmentation is written out. This would not be the case in a production version of the program.

Nearly finished segmenting bright regions...

This final image just shows the results of continuing to edit the relatively bright regions.

The important thing to notice here is how much the tool has simplified the process of creating a segmented image. It finds nearly all of the boundaries and accurately tracks them. The corrections that the user is making are more precise and more consistent than if the user was required to hand digitize every boundary of every region.

If you have questions, you can email them to:

Bill Langford

This file: http://www.nceas.ucsb.edu/~langfob/html/michelle.html
Last modified: Thu Jul 08 22:47:32 Pacific Daylight Time 2004