Mobilenetv2_CA Lightweight Object Detection Network in Autonomous Driving

Round 1

Reviewer 1 Report

High precision detection an real-time detection:

A well-written article, technically sound. Please expand your discussion referring to the computational efficiency of your technique in terms of computational resources, robustness, and "near real-time" latency.

Please, expand your discussion section addressing possible application of lightweight network target detection algorithms in areas beyond autonomous driving.

Minor editing.

Author Response

Dear reviewers, we have seriously responded to your review comments, and the details are in the attached PDF documents. Thank you again for your contribution.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this study authors combine Mobilenetv2 with 1) Cоrdinate attention mechanism, 2) PANet (for multi-scale fusion), and 3) Yolo detection heads. To construct a lightweight object detection neural network that can be used on autonomous driving platforms. They also propose a Mosaic image enhancement technology for a more correct learning procedure.

As a result, the proposed detector achieves impressive quality on the KITTI and the VOC datasets.

However, there are some changes that are required to improve the article:

●      In the introduction section, you mention different ways to accelerate the inference of the CNNs, but since your work is about fast and accurate object detection, you should also write something about object detectors there.

●      Please, clearly state your contributions at the end of the introduction section.

●      It is stated that Figure 2 shows the effects of data enhancement. In fact, it shows that a detector successfully captured small/obscured objects in difficult conditions. You should probably add images without data enhancement, to illustrate that without it the detector fails.

●      It seems that you use pre-trained MobileNetv2 as a backbone. Is that so? Please clarify in Section 3. B.

●      In Table 3 line break is unfortunate the version of the Mobilenet is on the new line “Mobilentv/[line break]/2”. Could you keep them on the same line?

●      In Figures 7-9, you used green bounding boxes for pedestrians and red for cars. Intuitively red color is associated with something wrong/incorrect, could you please change it?

●      It seems that the Discussion section is more of a Conclusion (maybe it should be renamed).

●      Future directions should be mentioned in the conclusion.

Also, there are two points that should be clarified:

●      Why did you choose Mobilenetv2 as a backbone? Why not Mobilenetv3? Or any other lightweight CNN?

●      Since you aim to provide a detector for autonomous driving platforms, computational performance is important. So the detectors should also be compared in terms of the number of operations and/or running time on mobile devices.

Finally, English editing and refinement may be a good idea. For instance, consider

●      Replacing citations like “The literature [#]...” with something more appropriate like “According to [#]...”, “The authors of [#]...” etc.

●      Correcting the grammar. For example “These methods of compressing the existing convolutional neural networks usually make the convolutional neural networks tend to be shallow.” You should either use “... make the convolutional neural networks shallow”, or “As a result of those methods, the convolutional neural networks tend to be shallow”. There are more grammatical issues.

●      Correcting typos both in words and in formulas. For example in Section 3B when you write “The initial learning rate is set to 10−3” you probably mean (10^-3).

Author Response

Dear reviewers, we have seriously responded to your review comments, and the details are in the attached PDF documents. Thank you again for your contribution.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper is quite nice and will need some minor changes in my opinion before it can be formally accepted. 

1. The abstract needs to be shortened and rewritten. Just focus on the results achieved in the abstract. 

2. Add a Conclusion section. Simply do not end the manuscript with a discussion section. 

3. Is the algorithm specifically tuned for the KITTI and VOC datasets? It would be good if the authors can provide the results of their implementation on other related datasets as well. 

4. The introduction section is too long. Simply put a table summarizing the results from the state-of-art. 

Author Response

Dear reviewers, we have seriously responded to your review comments, and the details are in the attached PDF documents. Thank you again for your contribution.

Author Response File: Author Response.pdf