Traffic Sign Detection Based on Lightweight Multiscale Feature Fusion Network

Round 1

Reviewer 1 Report

There are several broken sentences.  For example,
1) Zhe et al [9] created a traffic sign database 55 TT100K based on Tencent Street View and labeled every Xie et al [10] proposed a two-56 stage cascaded convolutional neural network framework, which can effectively improve 57 the classification accuracy of traffic signs.
2) The base structure of this network is Bottleneck, and the core idea is to reduce the computational complexity by deep separable convolution, thus replacing the standard convolution, in addition to In addition, a lightweight attention module is added to enhance the learning capability of the network.
3) In this paper, we use the summation operation for feature fusion, and the consistency of channel number is achieved by the convolution kernel of The green arrow in the figure represents the bilinear interpolation, and the size of the feature map after bilinear interpolation becomes twice of the previous one, and the transformed feature map can be fused with the underlying feature map.

The authors are suggested to proofread the manuscript to fix the broken sentence issues.

Page 10, the authors stated "Among them, a column is the original graph in TT100K dataset, b column is the detection result of Faster R-CNN, c column is the detection result of CornerNet, d column is the detection result of CenterNet, and e column is the detection result of MFHA-TSDR, the proposed algorithm in this paper."  However, MFA-TSDR does not seem to match the text.

2.1 Lightweighted Feature Extraction Networks: It is not clear why the base structure (Bottleneck) is considered as light weight.  It is not clear about the differences between BiFPN [25] and the proposed three-part multiscale feature fusion network.  The authors are suggested to include some experiments to show the effectiveness of the proposed feature fusion network compared with BiFPN.
2.2 Hybrid Attention Models: Some notations defined in the equations are not shown in Figure 5 and Figure 6 or not clearly defined.  It is not clear the channel and spatial attentions are proposed by the authors or are adopted by the authors.  Regardless, the authors should compare the attentions with the conventional attentions to see their effectiveness.
2.3 Detection Network: It is not clear about the differences between CenterNet [27] and the proposed detection network.  The authors are suggested to include some experiments to show the effectiveness of the proposed detection network compared with CenterNet.

The authors compared the proposed algorithm with several similar algorithms Faster R-CNN, CornerNet, and CenterNet.  More justification need to be added to explain why these methods are chosen as the compared methods.  It is not clear what are the fundamental differences between the proposed method and the other compared methods.  The authors stated the real-time performance issues and unreasonable allocation of computer resources.  However, they never compared the time complexity and resource allocation issues of different methods to show the effectiveness of the proposed method.  I would suggest that authors add more experiments to show the effectiveness of the proposed method in real-time performance and allocations of computer resources.  

The authors need to add more explanation on the classes in Table 2 and provide more analyses on why the proposed method works well on some classes and does not work well on other classes.

Author Response

Please see attached document.

Author Response File: Author Response.pdf

Reviewer 2 Report

I find the work acceptable for publication, provided that some minor issues are resolved. I.e., Some English corrections are required such as:

line 84, "combined" instead of "combed", line 107, correct the repetition "in addition to In addition", or, line 144, correct the two sentences "the convolution kernel of The green", ...

 

Author Response

Please see attached document.

Author Response File: Author Response.pdf

Reviewer 3 Report

Please see attached review comment document.

Comments for author File: Comments.pdf

Author Response

Please see attached response document.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Thanks for inviting me again, I think the authors have addressed my concerns, I am ok with the revision. 

Author Response

We are very grateful for the reviewer’s work. In all, the reviewer’s comments are quite helpful, which pointed out the deficiencies in our manuscript, and helped us with further improvement.

Once again, thank you very much for your comments and suggestions.