BEMF-Net: Semantic Segmentation of Large-Scale Point Clouds via Bilateral Neighbor Enhancement and Multi-Scale Fusion
, Sansheng Yang 3, Zhipeng Jiang 1, Jianjun Zhang 1, Shuhao Guo 1, Gaorui Li 2, Saishang Zhong 4, Zheng Liu 2,5,* and Zhong Xie 2Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIn the study, a new method for point cloud semantic segmentation is proposed. The superiority of the proposed method was tested on different data sets. The manuscript is generally well-designed and the explanations are sufficient. Some revision suggestions are given below.
1) In the introduction section, the disadvantages of working with KNN neighborhood are mentioned. But in your method you use KNN for feature extraction. How would you explain this situation?
2) There are many studies done on this subject. It is well known that color information improves point cloud semantic segmentation. Support the topic with references like below.
i) Atik, M.E.; Duran, Z. Selection of Relevant Geometric Features Using Filter-Based Algorithms for Point Cloud Semantic Segmentation. Electronics 2022, 11, 3310. https://doi.org/10.3390/electronics11203310
ii) Rim, B.; Lee, A.; Hong, M. Semantic Segmentation of Large-Scale Outdoor Point Clouds by Encoder–Decoder Shared MLPs with Multiple Losses. Remote Sens. 2021, 13, 3121. https://doi.org/10.3390/rs13163121
3) Discussion of the results should be improved.
4) Suggestions for future studies should be made in the conclusion.
Comments on the Quality of English LanguageMinor editing of English language required and make spell check.
Author Response
Our specific point-by-point comments can be found in the uploaded PDF file.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe following comments must be carefully revised.
1. Some important related work should be cited and analyzed, including “Spectrum interference-based two-level data augmentation method in deep learning for automatic modulation classification,” Neural Computing & Applications, vol. 33, pp. 7723-7745, 2020. Disparity-Based Multiscale Fusion Network for Transportation Detection. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(10), 18855-18863. Dual-Level Representation Enhancement on Characteristic and Context for Image-Text Retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 32(11), 8037-8050. Multi-Scale Fine-Grained Alignments for Image and Sentence Matching. IEEE Transactions on Multimedia, 2023, 25, 543-556. PUFA-GAN: A Frequency-Aware Generative Adversarial Network for 3D Point Cloud Upsampling. IEEE Transactions on Image Processing, 2022, 31, 7389-7402.
2. Specific details should be given in Fig. 1 and Fig. 2, such as the number of kernels and dimensions of feature maps.
3. Ablation studies should be performed to observe the effect of each module and each feature.
4. The convergence curve of loss function should be presented.
5. A more comprehensive experimental evaluation and comparison should be provided, especially considering the setting of various hyper-parameters.
6. Inference speed is encouraged to report.
7. The convergence process of the objective function should be demonstrated.
Comments on the Quality of English LanguageMinor editing of English language required.
Author Response
Our specific point-by-point comments can be found in the uploaded PDF file.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsSee attached file
Comments for author File: Comments.pdf
In general, with exception of some typos that do not tarnish the reading, the paper is well written.
Author Response
Our specific point-by-point comments can be found in the uploaded PDF file.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsAccept.
Author Response
Thank you for all your diligent efforts in reviewing our manuscript. Your valuable suggestions and professional opinions have profoundly influenced our research work. The recognition and support you have given us is very meaningful to us. We are deeply honored to have received your approval.
Moreover, your guidance has provided valuable insights for our research, and these suggestions will continue to play an important role in our future work. Once again, we appreciate your support and encouragement, and we anticipate more occasions to benefit from your guidance and assistance in the future.
Reviewer 3 Report
Comments and Suggestions for AuthorsThank you very much for the revised version, it is much clearer
I still find some minor inconsistency in the output dimensions in Figure 1 & 2:
- In Figure 1, E1 has size (N,8) and is given as in input to the second DBB. So after RS it shoiuld be of size (N/4,8). When passing it throught the second DBB, it should be of size (N/4, 16) and not (N/4,32) or there may be some MLP that are not explicitely presented.
- In Figure 2, when you input a (1,d) feature vector to the BLA, the outpus is (1,d/2). But in the description of the DBB, the second BLA gets an input of size (1,d/2) and has an output of size (1, d/2) which is not consistent.
In Table 4, there is a typo: U-Fuison -> U-Fusion.
Comments on the Quality of English LanguageEnglish language fine. Some minor typo issues detected.
Author Response
The point-by-point response can be found in the attachment PDF.
Author Response File: Author Response.pdf
