Next: Some poor quality images
Up: THE KTH-TIPS database
Previous: Image acquisition
Image cropping
To remove the background, and to be consistent with the experiments in
[4,5], we manually
cropped images to 200 x 200 pixels when possible. However, for
some samples (Brown bread and Cracker B) this was not possible at
large camera-target distances since the subject did not fill a
sufficiently large part of the image. In these cases, images were
instead cropped as follows:
- If possible, an ``equivalently sized'' rectangular region was
selected, with an aspect ratio as close as possible to 1.
``Equivalently sized'' implies that the same number of pixels should
be made available to the computer vision algorithm (e.g. material
classification algorithm) once early processing (e.g. filtering) has
been performed. In our work [3] the early
processing involved applying a filter bank, and removing pixels
which were not entirely covered by the support region of the filter
kernel. These pixels are located at the edges of the image patch.
In particular we used 41x41 filter kernels, so with a 200 x 200
patch, after filtering we were left with (200-40) x
(200-40) = 1602=25600 pixels which were input to the
classification algorithm. Therefore, we selected X x Y
patches such that (X-40) x (Y-40) was approximately 25600.
- If the largest possible X and Y did not satisfy the
``equivalently sized'' criterion above, we simply took the largest
possible rectangular region corresponding to the foreground texture.
We must emphasize that the ``equivalent size'' condition is dependent
on the employed image processing strategy and might very well be
poorly suited to your application.
Table 4 lists where these cropping strategies were
necessary. With Brown bread the texture round the edges of the slice
is somewhat different (denser) to that in the middle, so these edges
were also removed.
Table 4:
Images where it was not possible to extract 200 x 200 pixels foreground patches.
Material |
Scale |
Images |
Cropping strategy |
Brown bread |
6 |
All |
Equivalent size |
7 |
8,9 |
Equivalent size |
7 |
1,2,3,4,5,6,7 |
Largest possible |
8 and 9 |
All |
Largest possible |
Cracker B |
7 |
All |
Equivalent size |
8 |
1,2,3 |
Equivalent size |
8 |
4,5,6,7,8,9 |
Largest possible |
9 |
All |
Largest possible |
|
Additionally we would like to point out that with Orange peel it was
not always possible to extract 200 x 200 pixel foreground
patches either. However, with this material the CUReT database
exhibits similar problems; in the CUReT images some background is
present. Since one of our main objectives was to attempt to recognise
our samples using models trained on the CUReT database, we decided
against cropping the Orange Peel to a smaller size. It is, however,
undoubtedly a problem that the amount of background varies from scale
to scale, and our background was not quite as uniform as the CUReT
background.
Next: Some poor quality images
Up: THE KTH-TIPS database
Previous: Image acquisition