Author: Not specified Language: text
Description: Not specified Timestamp: 2018-01-21 19:46:52 +0000
View raw paste Reply
  1. \documentclass[10pt,twocolumn,letterpaper]{article}
  2.  
  3. \usepackage{cvpr}
  4. \usepackage{times}
  5. \usepackage{epsfig}
  6. \usepackage{graphicx}
  7. \usepackage{amsmath}
  8. \usepackage{amssymb}
  9. \usepackage{csquotes}
  10. \usepackage{multirow,graphicx,paralist}
  11. \usepackage{booktabs}  
  12. % Include other packages here, before hyperref.
  13.  
  14. % If you comment hyperref and then uncomment it, you should delete
  15. % egpaper.aux before re-running latex.  (Or just hit 'q' on the first latex
  16. % run, let it finish, and you should be clear).
  17. \usepackage[pagebackref=true,breaklinks=true,letterpaper=true,colorlinks,bookmarks=false]{hyperref}
  18.  
  19. %%%%%%%%% PAPER ID  - PLEASE UPDATE
  20. \def\cvprPaperID{2523} % *** Enter the CVPR Paper ID here
  21. \def\httilde{\mbox{\tt\raisebox{-.5ex}{\symbol{126}}}}
  22.  
  23. \begin{document}
  24.  
  25. %%%%%%%%% TITLE - PLEASE UPDATE
  26. \title{Revisiting Salient Object Detection: \\ Simultaneous Detection, Ranking, and Subitizing of Multiple Salient Objects}  % **** Enter the paper title here
  27.  
  28. \maketitle
  29. \thispagestyle{empty}
  30.  
  31. Many thanks to the reviewers for their insightful comments. We appreciate the reviewers acknowledgment of the many strengths of our paper, e.g. \enquote{very interesting and relevant difference to existing saliency systems}, \enquote{model design is interesting and reasonable} \enquote{novel perspective with good justification}, \enquote{can open up new directions in this field}, \enquote{new problem formulation}, \enquote{proposed network is effective}.
  32.  
  33. \vspace{10pt}
  34.  
  35. \noindent \textbf{Assigned Reviewer 1:}
  36. AR1 suggested to change \enquote*{\textit{detection}} to \enquote*{\textit{segmentation}} throughout the paper. We thanks AR1 for the suggestion and we agree with that; however, this problem is mostly treated as salient object detection in the previous works [7, 22, 24, 36, 37]. We will explain and fix some minor details recommended by AR1 in the final version of the paper. We believe the ground truth of the Pascal-S dataset [21] is not generated using the eye movements, rather, subjects were asked to select salient regions by clicking on them and the universal agreement among different subject was used to obtain the rank of each segment. We also would like to clear the confusion AR1 raised regarding the PASCAL-S dataset augmentation. For the saliency ranking and detection tasks, the PASCAL-S is augmented for relative salience; however, for the subitizing task, new ground-truth was provided that includes the number of salient objects in the image.
  37.  
  38. Rank order of a salient instance is obtained by averaging the degree of saliency within that instance mask. We can write the operations as follows:
  39. %In fact, we are in the process of optimizing the ranking mechanism to make it (best case scenario) as close as possible to higher cognitive  
  40. \begin{gather}
  41. \text{Rank} (\mathcal{S}_m^{T} (\delta)) = \frac{\sum_{i=1}^{\rho_\delta} \delta(x_i, y_i)}{\rho_\delta}
  42. \end{gather}
  43. where $\delta$ represents a particular instance of the final saliency map (${S}_m^{T}$) and $\rho_\delta$ denotes total numbers of pixels $\delta$ contains.
  44. We do not claim that the proposed way for obtaining the rank order is the best one possible. In fact, we expect that this work will inspire significant interest in exploring different ranking mechanisms.
  45.  
  46. Also AR1 asked to clarify Table 4. Yes, all the baselines number we reported from [34] and we did not include [6] in the comparison since they reported number on a different dataset (SOS-V2) whereas we evaluated on SOS-V1.  
  47.  
  48. Finally, AR1 also suggested to compare the object segmentation results on other datasets. We would like to reaffirm that training on other dataset is not suitable in our case (considering the concept of NRSS) since they do not provide agreement of multiple observers in the ground-truth. We can only evaluate our model on different datasets. Please note that other state-of-the-are methods are trained on a much larger dataset and optimized for a binary decision (ground truth provided in other benchmark). With that said, we have evaluated our model (trained on a subset of Pascal-S) on ECSSD dataset(ref) and report the number in Table~\ref{table:quan_ecssd}. In comparison, even though we are not absolute best but performance is very competitive with state-of-the-art methods.
  49.  
  50.  \begin{table}
  51.         \centering
  52.         \resizebox{0.2\textwidth}{!}{
  53.                 \begin{tabular}{c|cc}
  54.                         \specialrule{1.2pt}{1pt}{1pt}\
  55.                         \multirow{1}{*}{$\ast$}& $F_m$& AUC  \\
  56.                         \specialrule{1.2pt}{1pt}{1pt}\
  57.                        
  58.                         Amulet [] & & \\
  59.                         UCF [] &84.3 & 98.4\\
  60.                         \midrule
  61.                        
  62.                         \textbf{RSDNet-R} & 84.6 &97.7\\
  63.                        
  64.                         \specialrule{1.2pt}{1pt}{1pt}
  65.         \end{tabular}}
  66.         \caption{Quantitative comparison of methods including average F-measure and AUC on ECSSD dataset.}
  67.         \label{table:quan_ecssd}
  68.  \end{table}
  69.  
  70. \vspace{5pt}
  71.  
  72. {\small
  73. \bibliographystyle{ieee}
  74. \bibliography{rebuttal}
  75. }
  76.  
  77. \end{document}
  78.  
View raw paste Reply