FusionPillars: A 3D Object Detection Network with Cross-Fusion and Self-Fusion
Round 1
Reviewer 1 Report
The paper sounds good, but some serious questions as follow, 1. Loss functions of "regression loss, classification loss, and directional loss" are used to make objective appraisal on which corresponding modules in fitting procedure of your whole model, and why? 2.the "bold-faced letter" should be marked equally for all comparision methods (for example, marked in Table 4), and the shortcomings(0.3) between numbers (for example, 92.15 Vs. 92.45) should be unconver the reasons (or explanation) in experiment discussion section. 3.It is well that the experiment provide the quantitative comparision from table 1-4, but this section missing the fusion results of visual images for qualitative analysis, and it must be offered. 4."These two modules (SAS and PVC) enhance the feature expression ability of small objects", and "Cars,Pedestrains,Cyclists" are taken as targets for example, how about the smaller important object of road sign, etc.?
5.Conclusin section, what's weakness of the method and the next step?
[Major Revision]
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The manuscript contributes to the Remote Sensing journal, however is important to add some comments that I alreday did to the manuscript body.
In the manuscript there are a lot of words together, you need to separate.
In Methodology section in line 142 add Loss Function.
In line 364 you must to describe bbox and aos benchmark, like you describe bev and 3d benchmark.
In all manuscript write FusionPillar instead of fusionpillar and PointPillar instead of pointpillar.
In Figure 8, in Z axis please write speed sample.
Finally, you have to discuss your finding with other researchers (compare findings).
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Overall, the paper is good, organised, well written and the content is significant for the application of LiDAR in cars and robots. The authors propose a one-stage end-to-end network named FusionPillars to fuse multisensor data (namely LiDAR point cloud and camera images).
Some strengths related to both scientific content and form are:
- The paper is well-structured and documented.
- An overview on related works, some from recent years, is presented. It contains studies from specific literature about how different methods or techniques were used.
- The authors describe their original contribution related to the topic in discussion.
- The research methodology is well-defined and justified from a scientific point of view.
- The theoretical background is properly used.
- The research design contains experiments which are well described, and the results are presented accordingly and discussed.
- The conclusions are clearly defined.
- The references are in accordance with the topic of the paper.
Some corrections are needed:
- The abbreviations should be spelled when first appear in the text (even if there are known terms).
- Figure 1 is mentioned in Introduction, but it is not clear if it is original or not. The source should be indicated.
- Figure 2 is mentioned in Introduction, but it refers to Methodology. In general, the figures should be placed after they are mentioned in the text.
- The format of the paper should be improved.
- The use of English language should be revised.
- No future work is presented.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
