\documentclass{article} \usepackage[latin1]{inputenc} \usepackage{a4wide} \usepackage{graphicx} \usepackage{url} \usepackage{psfrag} \usepackage{pst-all} \usepackage{framed} \definecolor{dblue}{rgb}{0.2,0.2,0.6} \title{Answer to the reviews of paper 65: \\ Thick Line Segment Detection with Fast Directional Tracking} \author{P. Even*, P. Ngo, and B. Kerautret} \newcommand{\ok}{\newline {\color{blue} \underline{Corrected.}}} \newenvironment{answer} { \begin{framed} \begingroup\color{blue}% \textbf{Answer:} } {\endgroup\end{framed}} %\newenvironment{answer}{\\ \color{blue}\noindent\it {\underline{Answer:}}}% \newenvironment{todo}{\\ \color{red} \noindent\it {\underline{\textbf{TODO:} }}}% \begin{document} \maketitle We would like to thank the editors and reviewers for their work and for their constructive comments, questions and suggestions. Because the paper already reaches the 10 pages limit, and in order to avoid removal of possibly valuable contents for paper understanding, additional data are put on the github, that is referenced in the paper. A detailed list of the changes is given below with also some specific answers to raised questions. {\color{blue} \bf Our paper is attached to our answer and the proposed modified text is highlighted in blue color}. \section{Reviewer \#4} \begin{itemize} \item {\bf 1. Summary. In 3-5 sentences, describe the key ideas and experiments and their significance.} \begin{itemize} \item A straight line detector in gray-scale images is proposed in this paper. Proposed method is an extension of a previous work of the same author. Novelties brang in this version are 1) adaptive directional scan which provides extending the segment through apropriate direction, and 2) control of assigned thickness used in blurred segment recognition, where an initially assigned large thickness value is tuned regarding to the amount of augmentation at blurred thickness observed in subsequent iterations. \end{itemize} \item {\bf 2. Strengths. Consider the significance of key ideas, experimental validation, writing quality. Explain clearly why these aspects of the paper are valuable.} \begin{itemize} \item Experimenting on the synthesized images, it is shown that the proposed method brings performance improvement compared to the older version of the method. \item At the experiment done on the York Urban database, three literature works are outperformed in terms of two evalutation metrics, i.e. C and L/N. \item It is appreciated that the code of the work is published which encourages reproducibility. \item Writing quality of the paper is good. \end{itemize} \item {\bf 3. Weaknesses. Consider significance of key ideas, experiments, writing quality. Clearly explain why these are weak aspects of the paper, e.g. why a specific prior work has already demonstrated the key contributions, or why the experiments are insufficient to validate the claims.} \begin{itemize} \item Although performance improvement is achived comparejd to the previous version of the method, it is not compared with the latest state-of-the-art methods (e.g. [Almazan, E.J., Tal, R., Qian, Y. and Elder, J.H., 2017. Mcmlsd: A dynamic programming approach to line segment detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 2031-2039).]), but mainly with three literature works, latest of which was published in 2015. \\ \begin{answer} Thanks for pointing this out. We have added this reference to the paper. However, we do not consider it in the experiments, because contrarily to the other methods in which no parameter has to be set for comparisons, this method has several ranking level parameters which could largely influence achieved results. Moreover, the code is written in Matlab. To ensure a fair comparison of time performance, it would require a complete re-programming in C-like language, that could produce possible rewritting bias. \begin{todo} Ajouter la publi dans l'etat de l'art et un laius dans la partie expe pour dire qu'il n'est pas immediat de se comparer avec, et qu'on ne retient que les m\'ethodes les plus proches en terme d'architecture. DETAIL : voir si le protocole de comparaison propos\'e peut \^etre pris en compte dans la r\'eponse. \end{todo} \end{answer} Some notes are as belows: \begin{itemize} \item What is the reason to sum the blurred segment thickness by 1/2 in Eq. 5? \begin{answer} This value comes from a discrete geometry consideration linked to the space of all discrete lines, that could provide the observed digitization (called the pre-image). For instance, if we observe points on the same line, the measured width is 1 pixel. If we set the assigned thickness to this value, then we exclude possible one-line steps (well-known aliasing effect) for nearly horizontal lines. This value of 1/2 is assumed to bound all the possible expansions of the observed line. \\ The text was changed to precise the role of this half pixel margin. But of course, we have no space left to discuss all these discrete geometry considerations in the paper. \begin{todo} Done \end{todo} \end{answer} \item Could the authors present a couple of example for sythesized images that were used in the experiments and the performance of both versions of the method obtained on them ? \begin{answer} Due to page limitations, we could not add any figure nor respective performance result in the paper. However, an example of synthesized image is already available in the mentioned github, and we have completed the table with associated results. Accordingly to the measured standard deviations obtained on the whole set of 1000 randomly generated images, observed results on individual images can largely change. \begin{todo} Done. Second example ? \end{todo} \end{answer} \item What is understood from the paper is the performance results presented in Table 1 and 2 are obtained using unsupervised mode. Is it correct? Then, why the performance obtained by supervised mode was not also presented, despite Section 3.5 is allocated for explanation of it? \begin{answer} We confirm that the performance results are obtained using the unsupervised mode. However, as the unsupervised algorithm is based on the supervised multiple line detection algorithm, these results {\bf also reflect} the supervised detection performance. Furthermore, the supervised mode deals mainly with an interactive aspect, which was largely discussed in the former paper. In this context, there is a lack of available ground truth directly related to supervised line extraction. We guess that such ground truth would closely depend on application requirements. Thus, we rather concentrate here on the automatic detection novelties, assuming that achieved performance on all segments also impact the manual detection of one or several line segments. \begin{todo} V\'eriier que dans la partie supervis\'ee, on ne pr\'esente rien de sp\'ecifique, sinon, peut-\^etre qq chose \`enlever pour gagner de la place. \end{todo} \end{answer} \item It is not clearly mentioned and justified why the initial thickness values of 7 and 3 are used at the experiments of Table 1 and 2, respectively. Since one of the main motivation of superity of this work compared to the previous version is claimed as not using a fixed thickness input value, the justification of selection of these initial parameter values should be presented and performance related to selection of this parameter value should also be presented and discussed. \begin{answer} Experiments on synthetic images aim at assessing the improved performance beween formal and actual versions of the detector, including the thickness value estimation. The synthesized images are generated with a thickness varying from 2 up to 5 pixels. Therefore the initial assigned thickness is set to a greater value : 7. \\ The other detectors aim at providing thin lines and may reject too scattered image lines. To adapt to this behavior, we retrict the detection to thin lines using the initial assigned thickness to 3 pixels. \\ The text of the paper has been precised accordingly. \begin{todo} Done. \end{todo} \end{answer} \item I would like to see the performance of the previous version of the method [11] at the York Urban dataset, in Table 2 and Fig. 5. \begin{answer} Unfortunately, we have no more left space to extend Fig. 5, where images are already quite small. It would maybe be possible to add one line in the table, but the interest is weak in lack of the associated image. \\ We have added all the required informations in the github (https://github.com/evenp/FBSD), with the completed table (T, N and L values were already available, along with mean thickness W). We notice that achieved values have less meaning here, because as explained in the paper, the lines detected by the former method are more likely to incorporate spurious points, that artificially grow the width and length values. \begin{todo} Done. \end{todo} \end{answer} \item I would suggest to the authors to give the reference numbers for the methods ijn Table 2. If these numbers are obtained by the authors but not got from the papers of LSD, ED-Lines and CannyLines, then it may be better to refer to the methods with citing those papers at the caption of Table 2. \begin{answer} Thanks for this relevant suggestion. The caption is now completed. \begin{todo} Done. \end{todo} \end{answer} \end{itemize} \end{itemize} \item {\bf 4. Paper rating} \begin{itemize} \item Borderline \end{itemize} \item {\bf 5. Justification of rating. What are the most important factors in your rating? } \begin{itemize} \item A contribution of this work is it brings perfokkrmance improvement compared to the previous published version of the method. However, it is still not clear how the proposed method close to the state of the art performaonces reported by recent papers. \begin{answer} We would just like to insist on the fact that our method additionnally provides a measure of the line thickness without degrading other performance with respect to some other recent detectors. \end{answer} \end{itemize} \end{itemize} \section{Reviewer \#5} \begin{itemize} \item {\bf 1. Summary. In 3-5 sentences, describe the key ideas and experiments and their significance.} \begin{itemize} \item This paper considers the problem of detecting straight lines in grey-scale images, that is an essential task in several Computer Vision applications. The proposed approach is based on adaptive directional scans and control of assigned thickness, and it is evaluated on both synthetic and real scenarios. \end{itemize} \item {\bf 2. Strengths. Consider the significance of key ideas, experimental validation, writing quality. Explain clearly why these aspects of the paper are valuable.} \begin{itemize} \item Experimental results report good performances. The authors provide both the code and an online demo. \end{itemize} \item {\bf 3. Weaknesses. Consider significance of key ideas, experiments, writing quality. Clearly explain why these are weak aspects of the paper, e.g. why a specific prior work has already demonstrated the key contributions, or why the experiments are insufficient to validate the claims.} \begin{itemize} \item As acknowledge by the authors, the method is sensitive to the initial conditions. \end{itemize} \item {\bf 4. Paper rating} \begin{itemize} \item Accept \end{itemize} \item {\bf 5. Justification of rating. What are the most important factors in your rating? } \begin{itemize} \item Experiments clearly demonstrate that this paper is a valid contribution. \end{itemize} \end{itemize} \begin{answer} Thank you for reviewing our paper. As mentioned in the conclusion, we understand the weekness of the proposed method related to the initial conditions, we try to overcome these limitations in future works. \end{answer} \end{document}