Expand discussion

thesis/chapters/chapter04/benchmarking.tex

@@ -18,7 +18,7 @@ \section{Benchmarking Method}\label{sec:results_benchmarking}
 It consists of 20.000 hand-drawn sketches obtained via crowd-sourcing, that are
 evenly divided into 250 categories. To speed up computations, 50 of those
 categories are chosen to derive precision-recall statistics. From each
-category, an image is randomly chosen as the query and the rest is used as
+category, an image is randomly selected as the query and the rest is used as
 positive results. In this case, both the query images and the database images
 are from the sketch domain, so the effectiveness of the retrieval process
 without preprocessing biases can be examined.

thesis/chapters/chapter05.tex

@@ -4,6 +4,21 @@ \chapter{Discussion}\label{ch:discussion}
 that clearly favor one approach over the other. This might be explained by the
 multitude of influencing factors, that can impact the performance.
 
+Based on the results displayed previously, one can conclude that for
+cross-domain retrieval, an approach based on Canny edge detection and local
+features performs better than global approaches. Of the similarity measures
+used to compare the images' signatures, histogram intersection provides the
+most consistent performance. 
+The advantage over the global descriptors, however, is not large and poorly
+chosen parameter values can lead to the global descriptors outperforming the
+local variants.
+
+When both the query image and the database images are sketches, the situation
+seems to be reversed. The precision and recall statistics for the global
+LUMA+MEAN pipeline show a slight advantage over the local LUMA+PMEAN variant.
+This is probably strongly influenced by the images used in the retrieval
+benchmark, as will be discussed below.
+
 \section{Structural Choices}
 
 The intention underlying this thesis was to perform an evaluation of the
@@ -21,20 +36,25 @@ \section{Structural Choices}
 established distance metrics and the TF-IDF weighting scheme, that has been
 successfully applied to information retrieval for some time.
 
-Based on the results displayed previously, one can conclude that an approach
-based on Canny edge detection and local features performs better than global
-approaches. The distance however, is not large and poorly chosen parameters can
-lead to the global descriptors exceeding the local variants. This is probably
-strongly influenced by the images involved in the retrieval process, as will be
-discussed below.
-
 \section{Parameter Choices}
 
 Since many of the processing steps are based on commonly-used algorithms,
 literature already presented reasonable starting values for the evaluation. As
 the experiments in section \ref{sec:results_parameters} show, the initial
-values already produce competitive results.
+values already produce competitive results. The best parameter values seem to
+strike a balance between losing information due to small resolution and
+becoming overly sensitive to noise or unrelated image background. For local
+sampling methods, a neighborhood size of $\frac{1}{3}$ of the image dimensions
+repeatedly performs best. A value of $\sigma=1.5$ for the Gaussian blur of the
+Canny edge detector appears to be suitable to extract the edges that correspond
+to a human sketch of the object or scene. The advantage of an angluar resultion
+larger than $N_{\theta}=12$ for the curvelet transform is probably limited by
+to the poor accuracy of hand-drawn sketches.
 
 \section{Benchmark Dataset Choices}
 
-TBD
+Assigning general validity to the results presented above would be unjustified,
+because some properties of the benchmark datasets must be taken into account as
+possible biases.
+
+normalization in 2nd dataset

thesis/chapters/chapter06.tex

@@ -2,6 +2,8 @@ \chapter{Conclusion}\label{ch:conclusion}
 
 Conclusion goes here\dots
 
+cross-domain distribution is effect of semantic gap
+
 \section{Future Work}
 
-TBD
+why are some categories so difficult?