Road Segmentation and Environment Labeling for Autonomous Vehicles
Round 1
Reviewer 1 Report
The authors propose a labeling method called "relational labeling," which combines moving and stationary objects.
The paper's overall structure and organization are good. The paper contains five sections: "1. Introduction ", "2. Proposed Method", "3. System Design", "4. Results", "5. Discussion", and "6. Conclusions". The introduction section can be further improved and better organized to include more related work/literature review and a good number of references regarding state-of-the-art segmentation/labeling approaches.
Figures and table descriptions can be clearer and more descriptive to make it easier for the reader to understand the results. For example, fig 7 is very low res and not clear.
The reviewer suggests further expanding in the conclusion section on how you plan to incorporate more sensors, make meaningful data clusters, and further explain the consequence of having the number of clusters as a hyper parameter. Similarly, in the results and discussion section, evaluation of the model against some state-of-art segmentation methodologies is suggested to support your novelty and performance claims better.
Author Response
Please see the attachments.
Author Response File: Author Response.docx
Reviewer 2 Report
This paper proposes a pipeline to automatically label the 3D point cloud data. The key architecture includes an encoder, a K-means cluster, a decoder and a Neural network for automated labeling. Experiment results using representative datasets are given. The authors may consider the following comments, which might improve the current draft.
Major issues:
1. Section 1, Section 2 and Section 3.1, 3.2 are unnecessarily long. Many of them are common sense.
2. Since the proposed method is inspired by [15], the authors should compare their method with the one in [15].
3. Please explain in detail the design rationales in Figs. 2, 3, 4. e.g., how to decide the parameters of the systems.
Minor issues:
1. Typos are everywhere. Please proofread the draft before re-submission.
2. Please redraw all the figures, with at least 300dpi and 8ft font sizes.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The work presented has novelty and offers an opportunity for further improving the methodology to merge multiple frame receivers from other sensors. My only recommendation is to include/discuss current methodologies and how they perform compared with yours. Moreover, how uncertainty could be assessed and how you will deal with this uncertainty.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
My major concern is about the efficiency of clustering algorithm. K-means, requiring iterative optimization, is computationally expensive. The time used for clustering should be reported. In addition, I will suggest the authors to see the online clustering algorithm proposed in Rethinking Semantic Segmentation: A Prototype View and probably compare with it.
The literature review is not comprehensive. The work "Towards a weakly supervised framework for 3d point cloud object detection and annotation" also studies weakly supervised 3d labeling. In addition, the 2D method "Group-wise learning for weakly supervised semantic segmentation" should also be included.
Training details are missed. How do you train the network, more specifically, the parts A, C, D in Fig. 1?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
My previous comments were addressed
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 2 Report
I believe the manuscript in its current form is acceptable. However, the authors may want to check again the formats of the Tables and Figures.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 4 Report
The revision has addressed all my concerns. I am happy to accept it.
Author Response
Please see the attachment
Author Response File: Author Response.docx