Sparse Self-Prompt-Guided Stereo Matching for Real-World Generalization

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript introduces a sparse self-prompt guided stereo matching network that can be generalized across real-world environments. In general, the work is interesting and well presented. I have only a few concerns, listed below.

(1) What is the computational efficiency of the proposed SSPGNet for stereo matching? It is important for mobile vision systems to realize real-time stereo matching and 3D reconstruction.

(2) It seems the major contribution lies in a sparse self-prompt mechanism that allows disparity estimation in a coarse-to-fine way. It would be interesting to discuss whether this idea can be applied in mono- or multi-view depth estimation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors present a method of calculating depth maps for stereo vision data that they claim performs better for real life scenarios than the conptemporary state-of-the-art methods. A strong point of the paper is the fact tat authors do not limit the evaluation to only available datasets, but they also present the performance on a real-life data gathered by a stereo setup. 

As for most papers dealing with ANNs the details about the networks used are not specified, which makes it impossible to replicate the experiments.

The remarks I have are the following:

In a few points in the paper the authors give a certain value of the parameter that is not explained at all and the reader does not know why this exact value/values were used.

Some acronyms are not explained.

in formula 14 - why is the sum range set to 3 instead of 4?

You mention "evaluation website" at some point without explanation.

The results in tables are probably expressed in percent. Please mark it correctly or explain.

Also, the performance of the method is limited - from Table 1 we read that it takes 0.7 second to process a pair of images. What are the resolutions of the input pairs and depth maps? What hardware was used?

In Figure 7 it is difficult to pick any differences.

In Fig. 8 it should probably read "dense"? Also the text that references the Figure 8 is wrong - there are no different methods.

In the supplementary material the discussion of errors is missing.

 

 

 

 

Author Response

 Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1- The abstract did not present any numerical values ​​for the results or for any approved metrics to demonstrate the efficiency of the proposed system.

2- Add a table of related works summarizing the most important tools used in each research paper, in addition to the advantages and disadvantages of each work, and what was omitted in this manuscript compared to them.

3- Adding a flowchart that clearly and simply presents the proposed work for each stage of the work

4- Comparing the results with recent research papers published between 2023 and 2026

5- Add digital charts of results to track progress.

6- Enlarge figures 13-14 for greater clarity

7-Rewrite the conclusions more smoothly and coherently, isolating the results in their own section and allowing for reference to them when writing the conclusions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

All comments done