Advanced Feature Learning on Point Clouds Using Multi-Resolution Features and Learnable Pooling
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe PointStack model proposed in this study combines multi-resolution features and learnable pooling, taking into account both the global structure and local shape structure of point features. The paper is well-written and achieves good performance in terms of accuracy on two types of tasks. After carefully reviewing this manuscript, I have the following suggestions:
1. In line 47, it is recommended to supplement relevant literature to support the impact caused by the sampling method and pooling function.
2. Multi-resolution feature deep learning is one of the core technologies in this research. Section 2.2 lacks discussion and literature support for the latest methods.
3. In Section 4.2, it is recommended to add visualizations of the classification results for comparison or qualitative analysis, making the effectiveness of the model more intuitive.
4. In Figure 3, it is suggested to include classification results of other networks for comparison, such as PointMLP.
5. In Table 1, it is recommended to add comparisons with results from the latest research methods.
6. After adopting multi-resolution feature learning and LP to focus on granularity information, compared to models like PointMLP, is the complexity and efficiency significantly affected?
7. In Section 4.6, using ScanObjectNN to further supplement ModelNet40 to explore the limitations of training samples on the model is a good idea. However, ScanObjectNN was originally used for training only 15 categories. Is it feasible to transfer it to the 40 categories of ModelNet40? How can you ensure that the decrease in accuracy is not caused by the increase in the number of categories?
8. It is recommended to include quantitative evaluation results in the Abstract.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsKey notes:
1. It is necessary to conduct additional experiments on semantic segmentation of the scene:
The article focuses on the problems of classification and segmentation of parts of an object (part segmentation). However, in order to fully evaluate the potential of the developed method, it is necessary to conduct additional experiments on semantic scene segmentation. The results of such experiments, reflected in the revised version of the article, will allow the authors to compare their method with other modern models.
2. Weak literature review (related work):
Although the paper contains comparisons with methods such as PointNet, PointNet++ and others, the relevant section needs to be expanded to include other works for a more detailed analysis of existing methods.
3. Analyze the possibility of generalization:
The current version of the paper is limited to experiments on relatively simple data sets. Authors should consider assessing the generalization ability of the developed method for different scenarios and more diverse data sets. In addition, it is necessary to add an analysis of the scalability of the proposed approach when working with large-scale point clouds.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis manuscript proposes PointStack, a new point cloud feature learning network that utilizes multi-resolution feature learning and learnable pooling (LP). PointStack can effectively represent both global and local contexts, allowing the network head to comprehend both the global structure and local shape details. In general, the manuscript is easy to understand and read. There are some places can be further improved in the manuscript:
1、 One major concern with this manuscript is that its contributions are not very convincing. The manuscript claims the major contributions among Lines 71-80. However, PointStack outperforms existing feature learning networks by only 1.5% and 1.9% in overall accuracy and class average accuracy, and the segmentation task only increased by 0.4%. Algorithm accuracy needs to be improved.
2、 PointStack has been compared and analyzed with existing feature learning networks in terms of classification and segmentation accuracy. How is the time efficiency of classification and segmentation compared to other algorithms? It is recommended to add this section.
3、 Section 4.2 Shape Classification it is recommended to add a classification result figure.
4、 Suggest discussing in a separate section, detailing the advantages and limitations of the proposed method in this manuscript.
5、 There are also some other issues with this manuscript, for example, Table 1 should appear in the main text first, and then the specific content of Table 1 should be presented. It is recommended to check the entire text.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have addressed my concerns. In my opinion, this manuscript can be accepted for publication.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsSuggest removing the black lines on the left or top of each object's point cloud in Figure 3.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
