Suppose we are interested in only the man-made objects in the TinyTown image. One thing that differentiates the man-made objects (cars, houses and the road) from the natural (trees) is that the man-made ones are made up of straight lines. Our hitherto discussed measure draws attention to contrast, no matter if it is caused by texture, jagged or straight lines. In the TinyTown image a scale and spatial information expansion shows that much of the total information comes from the jagged outline of the trees and their texture.
To capture only the man made objects in the scene we want to define our information measure on a scale-space expansion of lines in the image. We do that by preprocessing the image with line finders, as described at the end of the theory section.
The line finders are even long and narrow Gabor patches, of several angular orientations, with a response similar to that of the directionally sensitive cells in human visual cortex. The resolution parameter now refers to the length of these patches, and thus also the minimum line length required in the image to give a maximal response at that resolution and orientation. The total response is obtained as a RMS response over all directions. The resulting ``lininess'' image expansion can be seen in Fig. 11, along with the total contrast between the successive ``lininess'' images.
Figure 11: ``Lininess'' of the Tinytown image at different resolutions (left), and
total ``lininess'' information at each scale (right).
Figure 12: Scale and spatial distribution of the ``lininess'' information in
the TinyTown image.
In Figure 12 we see the effects of computing the contrast between the successive ``lininess'' distributions. This measure is still dependent on contrast, but it suppresses contrast not related to lines.
Although not very suitable for popping out complete whole objects in a complex image like TinyTown, the measure draws attention to areas with lines in the different scale lengths.
In the first three scales various features of the cars give the strongest responses. First comes the outlines of the high contrast white car, and the specularities on the black car turning off the road in the middle of the street. Later responses come from the whole bodies of the cars, as well as various areas of the ground being enclosed by them. (As mentioned earlier the measure is oblique to the object anti-object issue.)
The last two images have their strongest responses from the road. At 320 pixel line length we see the unobscured upper right and lower left lanes. At 640 pixel we get a strong response from the right edge of the road, presumably because the dark colored cars on the right side gives less of a disturbance to the road edge than the white on the left side. (The average intensity difference between the road and the ground next to it is only 30 on a 256 grey level digitized image, while it is over 200 between the road and the white car.)