Saliency Map Estimation Using a Pixel-Pairwise-Based Unsupervised Markov Random Field Model

Round 1

Reviewer 1 Report

Saliency Map Estimation Using a Pixel Pairwise Based Unsupervised Markov Random Field Model

 

by Max Mignotte

 

January 2023

 

The paper deals with the problem of unsupervised saliency map estimation. The proposed unsupervised method is based on new modelling by Hidden Markov Fields (HMFs). Parameters are estimated with original “Iterative Conditional Estimation” (ICE) algorithm, and the segmentation is performed with Bayesian classification “Maximum Posterior Marginal” (MPM) and parameter estimation with Iterative Conditional Estimation (ICE). The main contribution is to define the observation field with a pairwise pixel modelling allowing taking into account the non-local pairwise pixel 95 interactions (NLPPIs). The whole method is very sophisticated and calls on different advanced tools. The presented experiments are very convincing and produce quite competitive results compared to the state of the art.

.

I recommend the publication of the paper after some minor remarks and suggestions below have been taken into account.

 

Minor changes

Page 3, Line 102: Maximum likelihood and Least Squares are two different general parameter estimation principles; setting them together, without more specification, appears as somewhat confusing;

- P3, L126-127: delete one “then”;

- P4L147: for each pixel s;

- P4L151: “region 2 of size » … what does “2” refer to?;

- P4L158-162: the phrase “Since … for Vs” is maybe too long, which makes it somewhat confused

- P6L177: say a word on how the weighting factor is determined;

- P6L184-185: you validated it by some statistical procedure, or just “visually” noticed? Have you tried other densities shapes? According to lines 194-195, there exist other possibilities;

- P6L184-185: ”for a given s” or ” for given s,t ”? ;

- P6L197: same question as above: have you tried other shapes?;

- P7 Fig2: as you have the ground truth, have you also used it to estimate these densities? This would give some indications of the effectiveness of ICE; - P8L216-217 are not clear… (7) simply results from the classic conditioning rule, there is no need to justify it; - P8L225: typo () before “see Fig.1”; - P8L239-242: You could cite here the paper “Y. Delignon, A. Marzouki, and W. Pieczynski, Estimation of Generalized Mixture and Its Application in Image Segmentation, IEEE Trans. on Image Processing, Vol. 6, No. 10, pp. 1364-1375, 1997.”, which shows an excellent behaviour of ICE in more complex context. Indeed, not only shapes of densities vary with the class, but in addition they are not known, and the proposed “generalized” ICE searches the shapes and estimate the parameters; - P8-9L249-251: “The fix point …. error [51]”: this is maybe true, but not sure and rigorous mathematical study of the asymptotic behaviour of ICE iterations is not easy (at the best of my knowledge, this was only showed in the very simple case discussed in [52]). Rather say that it was demonstrated in the simple case [52], and that numerous past experiments show its good behaviour. Not that the same is true for EM, despite claims that “EM converges” in numerous papers; - P10L281: the sequence being stochastic, how do you measure this third point? ; - P10L287: how the temperature of 0.15 was fixed?; - P10L297-300: The phrase “In addition, …to estimate” is maybe useless, unless you justify why the same difficulties do not exist for MPM; - P10L306: such a formulation is somewhat awkward as both MAP and MPM are “global” classifiers in that all observations are used to classify Xs. They just correspond to two different loss functions. Moreover, Xs or X(s,t)?; - P10L309: why (12) is cited here is not clear; - P11L310: , Xs or X(s,t)?; - P11: The phrase lines 318-323 is maybe somewhat long; - P17L467: why in “least square sense”?

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

authors  presents a Bayesian statistical approach to the saliency map estimation problem. it is a interesting topic. now I have some questions disscussed with authors.

1) According to authors, the highlights of this paper  lies in the use of a a pixel pairwise modeling and a likelihood model based on a parametric mixture of two different class-conditional likelihood distributions whose parameters are adaptively and previously estimated for each image. this will requires a large amount of computation， I suggest that  some comparison  based on computation cost should be added in the experiments。

2)  the  parameters in the proposed algorithm  also should be listed with compared algorithms。

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

This manuscript in this current form still has some problems as follows: 

1.Abstract: The abstract can be said more to highlight their own contribution.

2. The paper needs to add a vision of future work.

3. The English expression of the paper needs to be optimized.

4. In terms of literature research, it is suggested to add the description for the following work: “Symmetric implicational algorithm derived from intuitionistic fuzzy entropy”, Iranian Journal of Fuzzy Systems.

5. The format of the whole paper is not very uniform

6.When there is noise in Saliency Map, how to deal with it?

7. Section 2.4.3:   1. 2. 3. ------> 1), 2), 3)

8. What are the theoretical advantages of the proposed method?

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

It is a revised version. My comments have been addressed. It can be accepted now.