expeAuto.tex

The next experiments aim at evaluating the performance of the new
detector with respect to the previous one on a test set composed of a
selection of 20 real images.
One of them is displayed on \RefFig{fig:auto}.
Compared measures $M$ are the execution time $T$, the amount $N$ of detected
blurred segments, their mean length $L$ and their mean width $W$.
For the sake of objectivity, these results are also compared to the same
measurements made on the 20 images data base used for the CannyLine line
segment detector \cite{LuAl15}.
\RefTab{tab:auto} gives the achieved results.
\begin{figure}[h]
%\center
  \begin{tabular}{
      c@{\hspace{0.1cm}}c@{\hspace{0.1cm}}c@{\hspace{0.1cm}}}
    \includegraphics[width=0.32\textwidth]{Expe_auto/buro.png} &
%    \includegraphics[width=0.32\textwidth]{Expe_auto/coloredOld.png} &
%    \includegraphics[width=0.32\textwidth]{Expe_auto/coloredNew.png} \\
    \includegraphics[width=0.32\textwidth]{Expe_auto/bsOld.png} &
    \includegraphics[width=0.32\textwidth]{Expe_auto/bsNew.png} \\
    \includegraphics[width=0.22\textwidth]{Expe_auto/buroDetail.png} &
    \includegraphics[width=0.22\textwidth]{Expe_auto/dssDetailOld.png} &
    \includegraphics[width=0.22\textwidth]{Expe_auto/dssDetailNew.png}
    \begin{picture}(1,1)
      {\color{red}{
        \put(-19.5,31){\framebox(28,9)}
        \put(-5.5,31){\vector(-2,-1){20}}
        \put(-133.5,31){\framebox(28,9)}
        \put(-117.5,31){\vector(-2,-1){20}}
        \put(-247.5,31){\framebox(28,9)}
        \put(-231.5,31){\vector(-2,-1){20}}
      }}
      {\color{dwhite}{
        \put(-291,32.5){\circle*{8}}
        \put(-177,32.5){\circle*{8}}
        \put(-63,32.5){\circle*{8}}
        \put(-302,4.5){\circle*{8}}
        \put(-188,4.5){\circle*{8}}
        \put(-75,4.5){\circle*{8}}
      }}
      \put(-293.5,30){a}
      \put(-179.5,30){b}
      \put(-65.5,30){c}
      \put(-305,2){d}
      \put(-191,2){e}
      \put(-77.5,2){f}
    \end{picture}
  \end{tabular}
  \caption{Automatic detection on real images:
           an input image (a), the segments found by the old detector (b)
           and those found by the new detector (c), and a detail of the
           image (d) and the enclosing digital segments for both old (e)
           and new (f) detectors.}
  \label{fig:auto}
\end{figure}
\begin{table}
\centering
\input{Expe_auto/perfTable}
\caption{Measured performance of both detectors on standard images.
         $M_{old}$ (resp. $M_{new}$) denotes the measure obtained with
         the previous (resp. new) detector.}
\label{tab:auto}
\end{table}

The new detector is faster and finds more edges than the previous one.
Details of \RefFig{fig:auto} d) and e) illustrate achieved
accuracy improvements.
Output segments are thinner but also shorter.
The control of the assigned width to fit detected segment width
has the side effect of blocking the segment expansion when remote parts
are noisier.
Found edges are thus more fragmented.
The relevance of this behavior depends strongly on application requirements.
Therefore the control of the assigned width is left as an option, the user
can let or cancel it.
%In both case, it could be useful to combine the detector with a tool
%to merge aligned segments.

%Although these observations in unsupervised context should be reproduced
%in supervised context, similar experiments require an application context
%to dispose of a ground truth and of real users to assess the detector
%relevance through ergonomy evaluations.