Compressive Bidirectional Reflection Distribution Function-Based Feature Extraction Method for Camouflaged Object Segmentation

Round 1

Reviewer 1 Report

This article deals with the problem of object segmentation in the specific context of camouflage, where differences between object and background are minimized and naïve segmentation algorithms might be especially prone to errors. The authors’ proposal is to introduce the Cook-Torrance BRDF model and multiple illumination angles to aid in the segmentation procedure by identifying differences in the angular responses of the material and background. The authors have clearly stated the domain, goals, and methodology of the paper, and have validated their model with some simple test experiments under varying illumination angles, in comparison to more basic segmentation procedures.

In general, the article is clear, although it would benefit from substantial editing. Unfortunately, I have big concerns regarding its inadequate framing. The reference list is incomplete and insufficiently discussed, making it difficult to determine the degree of originality of the work or how it compares to other algorithms and proposals. The authors should reference up-to-date quality developments in camouflaged object segmentation, such as J. Yan et al., IEEE Access 9, 43290-43300 (2021). They also do not acknowledge other BRDF-based segmentation algorithms (see O. Wang et al., 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2805-2811; as well as the similarly named US 2015/0012226A1 Patent, registered in Jan. 8, 2015). Finally, they also do not clarify in which cases could their method, which is based on angle-dependent illumination, aid in camouflage reconnaissance. Although they acknowledge the limitations of the study in the last paragraph of the conclusions, they do not provide examples of real applications of the model.

Therefore, I believe the work to be incomplete in its relation to the current literature and developments in the field, which is particularly important in a work which, in addition to optics, is concerned with the computer-science aspect of imaging.

Other minor issues that I could find:

1. The introduction is very redundant, particularly from line 65 (page 2) onwards.

2. Figure 1 does not illustrate any physical model, but rather the geometry of the problem.

3. The Cook-Torrance model is not adequately described. In particular, the Gaussian parameters shown in Figure 2 are not explicitly shown in Equation 2 or elsewhere, nor are the other parameters of the model described.

4. Figures 2 & 4: the BRDF should not be called simply “Reflectance” and its unit should be sr-1.

5. Section 3 should be called “Model construction”, or an equivalent alternative.

6. The agreement between the CT model and its Chebyshev reconstruction in Figure 4 is not very good. It seems that both curves are offset by a scale factor or a constant, but the authors do not mention this. Furthermore, how many coefficients were used in the expansion? The authors should discuss the compromise between the number of terms, precision, and the sensitivity of the model.

7. What are the characteristics of the illumination source in Figure 5? Is it a broad- or narrow-band source? What is the divergence of the beam?

8. Equation 16: sin(pi(2j-2)/180), rather than sin(2*j-2).

9.Figure 7: the last two images of Group 3 are incorrect; the authors have duplicated the ones above them, which correspond to Group 2.

10. Page 11, line 182: the authors say “Guo et al. [8]”, but Ref. [8] corresponds to Goccia et al., with Guo et al. begin numbered [12]. In any case, I consider it inappropriate to attribute Methods 2 & 3 to any particular group, as they are basically one- and multi-dimensional grey-level segmentation procedures.

11. Figure 8: the red labels for the incidence angles are poorly seen, especially when printing in black & white.

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

In this work, the authors have presented a compression bidirectional reflection distribution function-based feature extraction method to achieve effective camouflaged object segmentation. The above paper is logically written and provides the necessary knowledge of the background of multiple conventional methods, the theory of the proposed method, and experiments to verify the accuracy and running speed. However, some specific problems should be improved. Therefore, I would like to make some specific comments as follows.

1. As mentioned in the paper, running speed is one of the remarkable advantages over conventional methods. Therefore, I suggest for the authors supplement more information about the hardware and software mentioned in lines 159 to 160.

For example:

a) What software did the authors use to implement the codes?

b) How did the authors use the mentioned PC to implement the codes and calculate the running time, one CPU core or multiple CPU cores?

2. This paper is aimed to achieve effective segmentation for camouflaged objects. Although the grey value of the object and background are very close for some images in Figure 6, the authors should better point out what specific kinds of camouflaged objects and backgrounds they used in the experiment for a better understanding.

3. Although it can be found that the coefficients of Chebyshev polynomials tend to zero after the first 20 polynomials, the authors should better point out the reason to reconstruct the Cook-Torrance model based on the first 20 coefficients in Figure 4.

4. Some minor revisions are also suggested.

For example:

a) In line 160, the authors have mentioned “The size of the single channel grey images to be segmented is 250 × 250.” I suggest for the authors supplement the unit of the image size.

b) In lines 203 to 204, the authors have mentioned “Compared with Method 2, the probability of the proposed method is improved 203 by 14.69%. and the false alarm rate is reduced by 35.78%.” It seems that it should be a comma after 14.69%.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This paper proposes a new object segmentation method with high precision and running speed. However, the application of this method is limited. My detailed comments are as follows:

(1) The conditions of this method are high. Can the author give the specific application of this method in introduction section?

(2) The background in the experiment is simple and uniform. It is necessary to add the experimental data when the background is uneven and there are multiple targets, so as to enhance the credibility of the paper.

(3) A detailed comparison of performance between the proposed method and segmentation method based on gray image is recommended.

(4) In line 106, the expressions of D and G should be given, and the corresponding parameters should be given in equation (2).

(5) Is the roughness σ the same as the root mean square slope of facets?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I thank the authors for their extensive rewriting of many parts of the article. In order to be as concise as possible, I proceed to list the remaining issues that I have detected:

1. The patent that I referred to is the following one: S Skaff, Material classification using brdf slices, US Patent App. 14/092,492, 2015. It can be found on Google Scholar through https://patentimages.storage.googleapis.com/94/69/43/8c551dd1c77aa2/US20150012226A1.pdf. The same author seems to hold many other related patents.
2. Regarding redundancies in the introduction (response number 1), I think the authors misunderstood what I meant. I did not mean that the paragraph from lines 65-80 was redundant itself, but rather that the last three paragraphs of the introduction (lines 65-80, 81-93, and 94-99) seemed to contain basically the same information in the same order, thus sounding repetitive. I would appreciate it if the authors made some changes in this direction, although I apologize for not being sufficiently clear the first time.
3. Response 3: you forgot to define alpha.
4. Response 6: I still don’t understand why there is an offset between the theory and its reconstruction. I though the curves were shifted to distinguish them, but it seems that negative reflectance values were intrinsically part of the reconstructed result. How do you get these negative values and why don’t you add a constant parameter F_0 to Equation (2) to correct this? I understand that the methodology relies only on relative reflectance changes, but I’m puzzled by it.
5. Response 8: there’s still a pi factor missing in Equation (19), which reads sin((2j-2)/180), instead of sin(pi(2j-2)/180).
6. Response 9: the authors have corrected the mistake I pointed out, but I have now noticed that the figure has another error (Group 4: 3D, 5D, 7D, are duplicated images of the ones of Group 3: 3D, 5D, 7D). This was not as obvious as the first one, so I didn’t catch it the first time.

I consider the article worthy of publication if the authors correct these minor remaining issues.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

It has been revised according to comments, and the quality of manuscript has been greatly improved.

Author Response

Thank you for your comments and suggestions to us in the first and second round revision.