Three-Dimensional Point Cloud Object Detection Based on Feature Fusion and Enhancement

Round 1

Reviewer 1 Report

The article is well-structured and addressing specific aspects of the research. This organization makes it easy for readers to follow the progression of ideas from background information to proposed solutions and experimental results.

The background section provides a comprehensive overview of the current state of research in 3D point cloud object detection, highlighting the significance of the problem and the challenges faced. This sets a strong foundation for understanding the need for the proposed solutions.

Authors given detailed explanation of the proposed network architecture, including the various modules and their functionalities. This level of detail is crucial for readers to grasp the novelty and effectiveness of the proposed approach.

Proposed work well describes the experiments conducted using the KITTI dataset and provides an analysis of the results. This empirical validation adds credibility to the proposed solutions and demonstrates their efficacy in improving object detection accuracy.

The article effectively breaks down complex concepts related to point cloud processing and deep learning into understandable terms, making the research accessible to readers with varying levels of expertise in the field. The conclusion and future directions section offers valuable insights into potential areas for further research and development, demonstrating a forward-thinking approach to addressing ongoing challenges in 3D object detection.

While the proposed solutions show promising improvements, there could be further exploration into addressing feature insufficiency, particularly for objects with extremely sparse point characteristics. This could involve investigating novel techniques or algorithms specifically tailored to handle such cases. However, some minor corrections would improve the manuscript quality.

One suggestion for improvement could be to delve deeper into how the proposed solutions integrate information from different viewpoints, such as BEV features, RGB image features, and raw point cloud features. Providing more insights into the fusion process could enhance the understanding of the network's capabilities.

Author Response

The response is included in "Response letter1.pdf".

Author Response File: Author Response.pdf

Reviewer 2 Report

Hi authors, your manuscript did propose a promising approach/framework while the results were not very well presented. I hope you can further reinforce this part to make the article qualify for publication. Specifically:

1.          Page 1, line 17: experiments (typo)

2.          Page 3, Figure 1 cannot really be observed or reflect any information – it is too dark.

3.          Page 6, line 216-217: “Therefore, SCL method would be introduced…” – I don’t see how it is shown in the upper right of Figure 2.

4.          Page 8, line 255-256 and Eq.2: can the author please address how the alpha was selected to balance d1 & d2.

5.          Page 8, line 276-282: these two paragraph should logically be merged. Even better, they should also merge into the followed one (283-287). Also, the “Figure ??” in the text were confusing readers.

6.          Caption in Figure 6 should be corrected to match the sequence (a) (b) (c). Also, it seems to me Figure 6(a) was literally irrelevant to (b) and (c), wasn’t it?

7.          I do believe the authors need to rephrase section 3.4 as well as recheck the presentations of Figure 7 and 8. They were sort of confusing.

8.          Similarly, section 3.5 should be rephrased, Table 2, 3, 4 should be re-presented. The text “Table 3.2, 3.3” in line 386 should be rectified. Authors may need to address how the values in the tables were obtained. Suppose with the presented results in Table 2 and 3, the proposed approach is just slightly better than Part-A2 rather than outperforming.

Author Response

The response is included in "Response letter2.pdf".

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper proposes the fusion of multiscale datasets composed of maps made with Deep Layer Aggregation (DLA) and feature fusion module for BEV and the 'use of a point completion network that reinforces positional features. Supervised contrastive learning is applied to improve segmentation results.

The manuscript is well written of considerable interest to readers and has innovative features in the field. 

The English needs to be revised because it has formal errors.

Author Response

The response is included in "Response letter3.pdf".

Author Response File: Author Response.pdf

Reviewer 4 Report

This paper proposes a new solution that combine spatial-semantic feature aggregation with auxiliary feature enhancement, to address some of the challenges in 3D object detection. This research topic is very interesting in many fields of application, as for example the autonomous driving.

The paper is well written and structured, and only needs minor adjustments. In the attached pdf I inserted a few comments to consider. 

One recommendation: I noticed many typos in the text. Some of them I have pointed out, but I advise the authors to carefully check the text.

Comments for author File: Comments.pdf

English language is fine. I just highlighted an unclear sentence to be rewritten.

Author Response

The response is included in "Response letter4.pdf".

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Hi authors, thank you for the efforts in preparing the revision of this manuscript. It is obvious that most issues occurred in previous version have been well taken care. I have no further questions regarding the current version manuscript.