Bayesian Edge Detector Using Deformable Directivity-Aware Sampling Window

Round 1

Reviewer 1 Report

Although the idea of the manuscript is of interests, I have a few suggestions as follows:
1. I suggest the introduction to be narrowed down in terms of particular domain/scope in which this particular application lies.
2. There are a few mistakes in the English language. Some sentences are not well structured, and their meaning is not well clear. Please check the language carefully.
3. The article should contain more recent references.

Author Response

1. I suggest the introduction to be narrowed down in terms of particular domain/scope in which this particular application lies.

 

Response:

We agree with the reviewer’s opinion, the introduction has been shortened in terms of particular domain/scope. Furthermore, to faithfully reflect this revision, we also modify the paper’s title to “Bayesian edge detector using deformable directivity-aware sampling window”

 

2. There are a few mistakes in the English language. Some sentences are not well structured, and their meaning is not well clear. Please check the language carefully.

 

Response:

English language has been extensively and thoroughly checked and revised.

 

3. The article should contain more recent references.

 

Response:

Regarding the problem of more recent references should be contained, lines 411~418 are newly added references 23, 24, 25. Their related descriptions are given in lines 324-344.

 

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposes an edge-detector named GestEdge that is used to output the perceptual edges. However, the manuscript is hard to follow. After carefully reading it several times, I think the topic is interesting and I’d like to give a chance to the authors for revising their manuscript. Hence, I suggest a major revision. My concerns are listed below.

1. Clarity issues:

• This paper should cautiously and clearly introduce the concepts, i.e., edge-detector, perceptual edges, directivity, and especially their correlations to the rotoscope and virtual reality.
• It seems that the topic of this manuscript is edge-detector. If so, the paper should focus on the topic of edge-detector. I suggest the authors cut down the discussion on the rotoscope in the 2nd paragraph. Similarly, cut down unnecessary discussion in other sections.
• The authors should also review the edge-detectors such as Papari and Grigorescu, which are the baseline methods employed in the result section. In addition, it is better to review those edge-detectors in a unique paragraph.
• The authors should clearly formulate the problem they targeted to resolve, and next, concisely describe it after having reviewed those edge-detectors.
• In the discussion section, I suggest the authors avoid discussing their future works in detail.
• I strongly suggest the authors provide the intermediate results attached in Figure 2. Those results may help the reader to better understand the proposed method.
• What are the correlations among the concepts, such as target pixel, perceptual pixel, and interesting pixel? In addition, what is the relationship between target pixels and perceptual edges? It seems the target of the proposed method is to detect perceptual edges, but target pixel can be left away from a perceptual edge.  

2. Experiments

• The latest baseline method employed was proposed in 2011, they cannot demonstrate the superiority of the proposed methods. Newly proposed methods should be included in the revised manuscript, e.g., the two methods listed below.

[1] Wei, Xing, Qingxiong Yang, and Yihong Gong. "Joint contour filtering." International Journal of Computer Vision 126, no. 11 (2018): 1245-1265. (edge map can be used)

[2] Liu, Feihong, Xiao Zhang, Hongyu Wang, and Jun Feng. "Context-Aware Superpixel and Bilateral Entropy—Image Coherence Induces Less Entropy." Entropy 22, no. 1 (2020): 20. (The boundaries of the hierarchical superpixels can be used.)

• How many images employed to derive the FOM values in Table 1? The authors should briefly discuss the dataset used, if possible, public datasets should be included. Specifically, it needs statistical analysis results to convince readers in making a judgment.
• I suggest the authors show the directivity map of a real image.

3. Minor issues

• In section 2, there should have cited the literature for all three postulations if they are not proposed by the authors. If they were proposed by the authors, they should be discussed in more detail.
• The authors should explain every term employed in their equation, such as S^{B_n}_{max}, et al. In the condition term of Eq.(4), there seems to lose parentheses.

Author Response

1. Clarity issues:

 

This paper should cautiously and clearly introduce the concepts, i.e., edge-detector, perceptual edges, directivity, and especially their correlations to the rotoscope and virtual reality.

It seems that the topic of this manuscript is edge-detector. If so, the paper should focus on the topic of edge-detector. I suggest the authors cut down the discussion on the rotoscope in the 2nd paragraph. Similarly, cut down unnecessary discussion in other sections.

The authors should also review the edge-detectors such as Papari and Grigorescu, which are the baseline methods employed in the result section. In addition, it is better to review those edge-detectors in a unique paragraph.

The authors should clearly formulate the problem they targeted to resolve, and next, concisely describe it after having reviewed those edge-detectors.

In the discussion section, I suggest the authors avoid discussing their future works in detail.

I strongly suggest the authors provide the intermediate results attached in Figure 2. Those results may help the reader to better understand the proposed method.

What are the correlations among the concepts, such as target pixel, perceptual pixel, and interesting pixel? In addition, what is the relationship between target pixels and perceptual edges? It seems the target of the proposed method is to detect perceptual edges, but target pixel can be left away from a perceptual edge. 

 

Response:

1. We agree with the reviewer’s opinion, the introduction has been shortened in terms of particular domain/scope. Furthermore, to faithfully reflect this revision, we also modify the paper’s title to “Bayesian edge detector using deformable directivity-aware sampling window”
2. We have removed the discussion on the rotoscope in the 2nd paragraph.
3. Lines 250~259 are added to explain those baseline methods (Papari and Grigorescu) employed in the result section.
4. To facilitate understanding the flowchart of Figure 2, we use Figure 3 to schematically illustrate how the sampling window M^t deforms as the directivity-aware scheme iterates Steps (b)–(d) until convergence. (lines 153~160, figure 3.)
5. In lines 79-85, we have explained the meaning of target pixel, perceptual pixel, and interesting pixel, as well as their relations.

 

2. Experiments

The latest baseline method employed was proposed in 2011, they cannot demonstrate the superiority of the proposed methods. Newly proposed methods should be included in the revised manuscript, e.g., the two methods listed below.

[1] Wei, Xing, Qingxiong Yang, and Yihong Gong. "Joint contour filtering." International Journal of Computer Vision 126, no. 11 (2018): 1245-1265. (edge map can be used)

[2] Liu, Feihong, Xiao Zhang, Hongyu Wang, and Jun Feng. "Context-Aware Superpixel and Bilateral Entropy—Image Coherence Induces Less Entropy." Entropy 22, no. 1 (2020): 20. (The boundaries of the hierarchical superpixels can be used.)

How many images employed to derive the FOM values in Table 1? The authors should briefly discuss the dataset used, if possible, public datasets should be included.

Specifically, it needs statistical analysis results to convince readers in making a judgment.

I suggest the authors show the directivity map of a real image.

 

Response:

1. Table 1 was deleted for concise reason.
2. Lines 324~344 are added to show comparison test of superpixels as candidate pixels using a natural image public database BSD300, and references 23 and 24 also newly added in response to reviewer’s opinion. Furthermore, lines 340~344 discuss the different output results of GestEdge when using Canny candidate pixels or using superpixels as candidate pixels.
3. The newly added Figure 3 shows the directivity map.

 

3. Minor issues

In section 2, there should have cited the literature for all three postulations if they are not proposed by the authors. If they were proposed by the authors, they should be discussed in more detail.

The authors should explain every term employed in their equation, such as S^{B_n}_{max}, et al. In the condition term of Eq.(4), there seems to lose parentheses.

 

Response:

1. Regarding the three postulations, we have adjusted the descriptions in lines 87~110, and readers are referred to references 11,12, and 13.  
2. We have checked every term employed in equations, making sure they have been clearly explained.  

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors resolved some of my concerns, however, this manuscript still has a lot of redundant contents that leave from the main topic. These contents would better be cut down, especially in the abstract, and other sections should be smoothed by the native speakers, especially the introduction.