210
E. Kagioulis et al.
Fig. 3. Average error along routes for different window sizes. Images are compared
using IDF matching and the lowest resolution (4
◦ per pixel).
be thought of as an SMW with a window the size of the entire memory. Within
the appropriate range of window sizes, SMW produces half the mean error and a
flatter distribution of error than PM. The lower mean error can be attributed to
correlation coefficient being able to deal with occlusions and generate a smoother
familiarity gradient compared to IDF.
Image Matching: Correlation coefficient matching is superior to IDF matching.
For smaller window sizes, both methods produce large errors for the reasons mentioned in the previous section. When pre-processing with blur, the IDF requires
a larger window size to converge to the same low level of error as CC. After window size 13, the error resulting from both approaches is equivalent. Combining
IDF with edge pre-processing yields high error signals even when the window
size is large which we believe is due to occlusion having more drastic effects
on edges than it does on pixels. Our analysis showed that, when the memory
window contains an almost fully occluded image, test images tend to strongly
match with it. This is because when there are very few edges, the IDF is small
even if the edges do not match at all. The match is so strong that the window
stays trapped in that area of the memory for the remainder of the route. As the
data show (Fig. 4) the IDF performs relatively well with edge pre-processing,
this issue can potentially be avoided by increasing the window size.
In contrast, correlation matching performs better when combined with edge
detection than it does with blur. This is because a lot of noise is removed by
the edge detector. In addition, the correlation is normalised by the standard
deviation of the images, so that it measures better how edges match regardless
of how many edges there are. This avoids the false matches with occluded (edgepoor) views we discussed above for the case of the IDF (Fig. 4).
E. Kagioulis et al.
Fig. 3. Average error along routes for different window sizes. Images are compared
using IDF matching and the lowest resolution (4
◦ per pixel).
be thought of as an SMW with a window the size of the entire memory. Within
the appropriate range of window sizes, SMW produces half the mean error and a
flatter distribution of error than PM. The lower mean error can be attributed to
correlation coefficient being able to deal with occlusions and generate a smoother
familiarity gradient compared to IDF.
Image Matching: Correlation coefficient matching is superior to IDF matching.
For smaller window sizes, both methods produce large errors for the reasons mentioned in the previous section. When pre-processing with blur, the IDF requires
a larger window size to converge to the same low level of error as CC. After window size 13, the error resulting from both approaches is equivalent. Combining
IDF with edge pre-processing yields high error signals even when the window
size is large which we believe is due to occlusion having more drastic effects
on edges than it does on pixels. Our analysis showed that, when the memory
window contains an almost fully occluded image, test images tend to strongly
match with it. This is because when there are very few edges, the IDF is small
even if the edges do not match at all. The match is so strong that the window
stays trapped in that area of the memory for the remainder of the route. As the
data show (Fig. 4) the IDF performs relatively well with edge pre-processing,
this issue can potentially be avoided by increasing the window size.
In contrast, correlation matching performs better when combined with edge
detection than it does with blur. This is because a lot of noise is removed by
the edge detector. In addition, the correlation is normalised by the standard
deviation of the images, so that it measures better how edges match regardless
of how many edges there are. This avoids the false matches with occluded (edgepoor) views we discussed above for the case of the IDF (Fig. 4).