Stereo SLAM in Dynamic Environments Using Semantic Segmentation
Round 1
Reviewer 1 Report
This paper presents a stereo-visual SLAM with dynamic object removal using deep learning-based semantic segmentation. Moving Object Detection (MOD) is important for visual SLAM due to its negative effect in large-scale outdoor environments, so this paper adequately covers the topics needed in SLAM field. However, in order to increase the academic value of this paper, I recommend some revisions as follows:
1. Since MOD-based visual SLAM has been studied for a long time, there are a lot of research papers covering the subject. The authors can contain more related papers in Chapter 2.
2. In the experiments, the authors choose three comparison algorithms; ORB-SLAM, OPENVslam, and DYNAslam. It would be much better if the authors explain why these algorithms are selected and why these algorithms are proper to evaluate the proposed SLAM framework.
English is good enough to present research results.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
My major concern involves the usage of semantic segmentation method ENet. The method is published in 2016 and thus out-of-dated considering the fast evolving of the field. I am wondering how the method will perform if using more advanced semantic segmentation methods.
Following the previous comments, recent semantic segmentation methods should be reviewed, such as, Exploring cross-image pixel contrast for semantic segmentation, Rethinking Semantic Segmentation: A Prototype View, Volumetric memory network for interactive medical image segmentation.
As stated, the method can only recognize 19 categories defined in Cityscapes, but how will the method deal with new categories appearing in the scene?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The evaluated article is a sample application of visual SLAM using the semantic segmentation neural network Enet. Visual SLAM has been addressed in the scientific literature for about 20 years, but this article brings innovation in the form of the use of the moving object detection (MOD) method proposed by the authors. The authors document their own contribution of improving localization accuracy by testing the proposed VSLAM algorithm on the KITTY dataset and comparing it to other VSLAM methods. I consider the graphs in figures 4, 5, 6 and table 2 with the quantification of errors and their reduction using the proposed VSLAM algorithm, which is relatively significant, to be the key results of the research.
I consider the conclusions resulting from the presented research to be correct, even though the experimental verification of VSLAM activity was done at a (probably) low movement speed the TurtleBot3 robot. References describe the topic treated in the article. I recommend publishing the article after minor formal adjustments and completion of answers to questions.
Comments:
1. Please unify the colour legend on Figures 4 and 5.
2. Please add the units on the vertical axes in the graphs in fig. 4.
Questions:
1. What was the speed of the TurtleBot3 robot during the experiment in chapter 4.3?
2. In line 258, the threshold value N=5 is set. How does change the success of moving objects detection when the threshold value is increased/decreased?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The manuscript is well-revised, so I recommend its publication.
Overall English is good, but, I recommend that the author's final proofreading before submitting the final manuscript.
