ALReg: Registration of 3D Point Clouds Using Active Learning
Round 1
Reviewer 1 Report
The paper presents a novel approach called ALReg for point cloud registration, which focuses on reducing the training time of deep learning networks by selectively choosing informative subsets of the point clouds. The authors show that their approach outperforms state-of-the-art methods such as DeepBBS, FMR, and DCP on the ModelNet40 and 7Scenes datasets, while using only a fraction of the training data and time needed to converge.
The paper is well-structured, and the authors provide a clear motivation for their work. The methodology is adequately explained, and the experiments are well-designed, providing a detailed analysis of the results. The comparison with other state-of-the-art methods and the experiments on different datasets add to the paper's strength.
Overall, the paper presents a promising approach for point cloud registration, which could potentially reduce the time and resources required to train deep learning networks. However, I believe that there are some minor issues that need to be addressed before publication. Please find them below:
· High image quality is essential, please provide better quality.
· As the paper presents important process, it is required that the discussion section should be presented in a clear and separate section to convince readers that the proposed method is more effective than existing studies.
· Please note that your manuscript requires English corrections.
The paper presents a novel approach called ALReg for point cloud registration, which focuses on reducing the training time of deep learning networks by selectively choosing informative subsets of the point clouds. The authors show that their approach outperforms state-of-the-art methods such as DeepBBS, FMR, and DCP on the ModelNet40 and 7Scenes datasets, while using only a fraction of the training data and time needed to converge.
The paper is well-structured, and the authors provide a clear motivation for their work. The methodology is adequately explained, and the experiments are well-designed, providing a detailed analysis of the results. The comparison with other state-of-the-art methods and the experiments on different datasets add to the paper's strength.
Overall, the paper presents a promising approach for point cloud registration, which could potentially reduce the time and resources required to train deep learning networks. However, I believe that there are some minor issues that need to be addressed before publication. Please find them below:
· High image quality is essential, please provide better quality.
· As the paper presents important process, it is required that the discussion section should be presented in a clear and separate section to convince readers that the proposed method is more effective than existing studies.
· Please note that your manuscript requires English corrections.
Author Response
Dear reviewer,
Thank you for your valuable comments. We addressed your feedback and revised our article. You can find detailed explanations about the changes below.
- High image quality is essential, please provide better quality.
We sincerely thank the reviewer for this feedback. We find out that especially Figure 2 and Figure 4 have low quality images for point clouds from 7Scenes. We changed these images with new ones.
- As the paper presents important process, it is required that the discussion section should be presented in a clear and separate section to convince readers that the proposed method is more effective than existing studies.
We added a Discussion section to the manuscript.
- Please note that your manuscript requires English corrections.
We corrected the language mistakes with the help of Grammarly.
Reviewer 2 Report
Please see the attached file.
Comments for author File: 
Comments.pdf
Author Response
Dear reviewer,
Thank you for your valuable comments. We addressed your feedback and revised our article. You can find detailed explanations about the changes below.
-In the abstract it is written that the method has been tested also on 3DMatch dataset. A brief comment of the related results can be useful.
The following sentence is added under the abstract: “The trained models are also tested on the 3DMatch dataset and better results are obtained than the original FMR training procedure.”
-Many acronyms, especially in section 2, are not clearly defined. Some example can be found in line 81 (ISP), line 83 (SVD), line 93 (SO(2)) and line 94 (RGM).
Definitions for the acronyms are added especially in Section 2.
-Even if the registration framework is represented by literature networks, some further details can be interesting. (example, the considered loss function)
The detailed explanation of the loss function are added to the manuscript.
-Subfigures of Figure 5 are quite small. In addition, the standard RMSE metric gives lower values for better results. From both graph seems that the proposed method performs worsen of both full and random sampling version of the registration algorithm. In the manuscript, specifically in lines 298-300 it is written that:” According to Figure 5, even their metric performances are worse, it is observed that for larger rotations on 3DMatch, both ALReg methods provides reduced errors.” Such statement clearly doesn’t match with graphs of figure 5.
We changed the size of the Figure 5. In the given sentence, a comparison to the full point cloud registration was aimed. The sentence is changed as “both ALReg methods provide reduced errors than the full point cloud training”.
-Table 1 is the last recalled in the text. Typically, tables are placed in the same order with respect to their analysis in the manuscript. So, either change the placement of table 1 or its in-text comment position.
Table order is changed.
-3Dmatch dataset has been tested also in the Gaussian noise case (table4) and sampled point cloud case (table 5). Why these results are not comment in the text?
The sentence is changed as “We also evaluated ALReg's robustness against Gaussian noise as well as incomplete point clouds and used 7Scenes and 3DMatch as testbeds”.
Reviewer 3 Report
Reviewer Comments
· The motivation of the work with deep learning needs further clarification?
· The main contribution and originality should be explained in more detail?
· What is your fitness function? what is the outcome of the proposed algorithm?
· The presentation of the results is not state-of-the-art. More simulation results and formal comparison of results are needed?
· Parameters setting of proposed method must be provided.
· Please separate "Conclusion" from this section. DISCUSSION and CONCLUSION will be two separate sections.
· Need add more recent works on proposed work. Some works should mention in this paper:
“A multi-granularity semisupervised active learning for point cloud semantic segmentation. Neural Computing and Applications. 2023 Apr 17:1-7.”
“Change detection of urban objects using 3D point clouds: A review. ISPRS Journal of Photogrammetry and Remote Sensing. 2023 Mar 1;197:228-55.”
“Point Cloud Registration via Heuristic Reward Reinforcement Learning. Stats. 2023 Feb 6;6(1):268-78.”
“A Point Cloud Data-Driven Pallet Pose Estimation Method Using an Active Binocular Vision Sensor. Sensors. 2023 Jan 20;23(3):1217.”
Minor editing of the English language required
Author Response
Dear reviewer,
Thank you for your valuable comments. We addressed your feedback and revised our article. You can find detailed explanations about the changes below.
-The motivation of the work with deep learning needs further clarification?
The following paragraph is added to the manuscript.
“Our main motivation is that, by using fewer point clouds or point cloud parts in the training phase of any point cloud registration network, a similar accuracy score could be obtained. However, it depends to the efficient selection of these training samples/parts. By the usage of ALReg, we aim the involve the most effective point cloud parts to the training procedure.”
-The main contribution and originality should be explained in more detail?
A detailed explanation is added under the “Subsection 2.3. Active Learning for Point Clouds”.
- What is your fitness function? what is the outcome of the proposed algorithm?
The details for the loss functions are added to the article. The outcome of the proposed algorithm is selecting fewer point clouds while training the given registration networks.
The presentation of the results is not state-of-the-art. More simulation results and formal comparison of results are needed?
Due to the lack of time in the review period, no new results could be obtained unfortunately.
-Parameters setting of proposed method must be provided.
More parameters and details are added under experiment section.
- Please separate "Conclusion" from this section. DISCUSSION and CONCLUSION will be two separate sections.
Conclusion and Discussion are separated.
- Need add more recent works on proposed work. Some works should mention in this paper:
“A multi-granularity semisupervised active learning for point cloud semantic segmentation. Neural Computing and Applications. 2023 Apr 17:1-7.”
“Change detection of urban objects using 3D point clouds: A review. ISPRS Journal of Photogrammetry and Remote Sensing. 2023 Mar 1;197:228-55.”
“Point Cloud Registration via Heuristic Reward Reinforcement Learning. Stats. 2023 Feb 6;6(1):268-78.”
“A Point Cloud Data-Driven Pallet Pose Estimation Method Using an Active Binocular Vision Sensor. Sensors. 2023 Jan 20;23(3):1217.”
The given recent works are added to the manuscript.
Reviewer 4 Report
It is really interesting concept, this could be a good target towards medical sciences. I have just one concern if you can simplify the language so that clinical people also can understand. They can bring this concept to the medical science. Well written with appropriate citation
Author Response
Dear reviewer,
Thank you for your valuable and fine comments. We addressed your feedback and revised out article by adding new details and an applying some verbal changes.
We hope that the methods will also be applied in the medical science community, since registration is also a popular task for medical imaging.
Best,
Yusuf.
Reviewer 5 Report
In this paper, Authors presents an application of active learning for the purpose of point cloud registration using various deep neural networks.
The manuscript structure is fine, everything seems to be explained and the results, as presented, confirm the author’s claims. The manuscript should of interest to the readers and it could be accepted as-is.
Author Response
Dear reviewer,
Thank you for your valuable and fine comments. We hope that our article will draw attention among both point cloud registration and active learning communities.
All the best,
Yusuf.
Round 2
Reviewer 2 Report
Most of my previous comments have been addressed. One points remains unsolved. In particular, the position of the tables must follow their comments in the text. Therefore, the actual Table 2 must be placed later than table 3,4 and 5 or, conversely, the discussion about the train times must be addressed before all the quantitative results evaluations.
Author Response
Dear reviewer,
Thank you for your positive and constructive feedbacks. We changed the order of the tables and made the "Train time of the studied methods" the last table in the manuscript.
Reviewer 3 Report
NONE
Author Response
Dear reviewer,
Thank you for positive feedback.
