Permalink
Browse files

Expand discussion

  • Loading branch information...
1 parent 0082b04 commit 5d744ca6edd0ff8f9cfd4552ef6066222b57f885 @weltenwort committed Oct 12, 2012
Showing with 25 additions and 11 deletions.
  1. +25 −11 thesis/chapters/chapter05.tex
  2. BIN thesis/thesis.pdf
@@ -75,15 +75,29 @@ \section{Benchmark Dataset Influences}
the lack of translation invariance of the global descriptor. While that
invariance would be expected to be an advantage of the local descriptors in
general, it can be a confounding factor in this case, which may contribute to
-the relatively poor results.
+the relatively poor results. Whether translation invariance beyond the
+slight fuzzyness introduced by the sampling method would be desirable at all,
+clearly depends on the images involved and the expectations of the user. If the
+user sketches a scene in order to find photographs with similar composition,
+disregarding the location of drawn features would be counterproductive.
-As shown in Figure~\ref{fig:results_distribution} there are a few query images
-in the cross-domain benchmark, for which the proposed solutions don't lead to
-good rankings. Similarly, the results of the intra-domain evaluation
-(Figure~\ref{fig:results_precision}) show, that there are significant
-differences in descriptor performance between different sketch categories.
-Reasons for that might be that the sketches or images contain too much
-distracting patterns or that the datasets contain several representations of
-the same object, that are visually not very similar. It also indicates, that
-one set of pipeline components and parameter values might only be suitable for
-a limited range of image types.
+So while the intra-domain dataset's normalization of the images seems to favor
+global descriptors, the cross-domain dataset mixes query intents and image
+types. Some sketches and images depict single objects with varying degrees of
+context and background, while others capture whole scenes or small objects
+within a larger composition. Differences in descriptor performance within such
+a diverse image set is to be expected. And indeed, as shown in
+Figure~\ref{fig:results_distribution} there are a few query images in the
+cross-domain benchmark, for which the proposed solutions consistently don't
+lead to good rankings.
+
+But even in the intra-domain evaluation (Figure~\ref{fig:results_precision})
+there are significant differences in descriptor performance between different
+sketch categories. Reasons for that might be that the sketches or images
+contain too much distracting patterns or that the datasets contain several
+representations of the same object, that are visually not very similar.
+
+The overall picture is, that each set of pipeline components and parameter
+values might only be suitable for a limited range of image types. When the
+system is designed for a more specific purpose, its performance can probably
+exceed the results shown above by a large margin.
View
Binary file not shown.

0 comments on commit 5d744ca

Please sign in to comment.