Traffic Sign Detection Method Based on Improved SSD

Round 1

Reviewer 1 Report

In this paper, the authors proposed an improved SSD model for traffic signs detection. I have the followings comments to be addressed by the authors

1- The authors replaced some of the 3x3 convolutions with 1x1 convolutions in the base network of the SSD. This requires training the modified VGG net on the Imagenet again, however, this has not been done. The training dataset they used (TT100K) is relatively small to train the VGG network. The authors should explain how they solved this issue.

2- The authors should provide the number of the parameters of the baseline SSD and the modified SSD (before and after parameters reduction).

3- The selection of the ranges of H, S, and V should be explained more.

4- The authors used a threshold to tell if the final connected component is a traffic sign or not. This threshold is based on the number of the pixels of the connected component of the potential traffic sign. However, since the distance between the vehicle and the traffic sign changes constantly during driving then the number of pixels will be changed. I think the authors did not explain how to deal with this situation.

5- The authors compared their results with PVANet but there is no reference to it.

6- The authors have mentioned many deep learning-based models for traffic sign detection. However, they did not compare their results with any of these models. It is difficult to tell if their model performs well or not. Thus, I recommend the authors to compare with the state-of-the-art models to see the real performance of their models.

7- The authors should visualize some of the detection results of their proposed model.

8- Not all configuration of the proposed model are provided. I think it is difficult to reproduce the results of the manuscript. Authors are recommended to either provide all details related to model training or post the code on GitHub or any other platform. 

Author Response

Please refer to "reply letter-reviewer1. Docx" for details.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper proposes a novel SSD network algorithm with increased speed / accuracy and lower computational complexity as compared to popular traffic sign detection methodologies. The manuscript is well-structured, providing an overview of popular sign detection methods and their weaknesses, as well as demonstrating the proposed methodology and its evaluation in detail. The use of several figures/tables is positive, as it makes the paper more comprehensible. The paper requires some revisions as follows:

[Abstract] “some false signs are often detected” should be rephrased to avoid misunderstandings. “false detection of traffic signs” could be used instead.

[Abstract] “The detection effect of small targets is improved, and the amount of calculation for the baseline SSD network is large”. “But” could be used, instead of “and” to highlight the complexity of the calculation.

[Abstract and other subsections] “Further then the color detection algorithm” please rephrase “further then” which can also be found in several subsections of the paper.

[Introduction] “are one of the key research areas”, “and similar false signs”, “by segmenting areas of interest and use”, “has become the mainstream”, “we improve based on the baseline SSD network structure and combines the color detection algorithm to deliver a lightweight and high-precision detection scheme,”, “are summarized as follow”, “this model is used to detect traffic signs in the image, and mark the location of the traffic sign in the image and the traffic sign category to which it belongs.” Please rephrase.

[Introduction] “Since the shapes of traffic signs are colorful triangles, circles, and rectangles” Please rephrase as not all signs have these shapes.

[2.2] “been widely used with its excellent performance” Please rephrase.

[2.2] “The above algorithm demonstrates improvement based on the existing target detection network and makes the high precision recognition of traffic signs possible, but most of them” Please rephrase as it is not clear where “them” refers to.

[3] In Figure 1 “Whether the maximum connected component meets the threshold” would be better if it was rephrased to form a question.

[3, 3.1, 3.2 and other subsections] Imperative is used instead of Passive voice in many cases throughout the paper. Some examples are: “Firstly, use”, “Secondly, use”, “Thirdly, use”, “set the red, green, and blue coordinates”, “Take the 6 layers” Please use passive voice.

[3.2] Please confirm the use of the phrase “indicator signs” as some signs seem to be mandatory instruction signs.

[3.2] “The statistical results show that the HSV color space range of the main colors of the three types of

traffic signs is shown in Table 1” Please rephrase.

[3.2] “tear and wear” Please use “wear and tear”.

[3.3] “Among them, the four-neighboring method” Consider removing “Among them”

[3.3] “If the connected component with a smaller number.. the set threshold “ Very long sentence, please rephrase.

[Figure 9] Sentences starting with “whether” should be stated as questions to me more understandable. In boxes where two steps are stated, these steps should be interconnected with one another to show that yes/no refers to both of them.

[4.1] “In formula (4), α > 0 and β are the gain and offset value, which respectively controls the brightness

and contrast of the image” Please rephrase.

[4.1] In formula (5), please adjust the font size.

[4.1] “In formula (6)(7) the test set” Please use plural.

[4.1] In table 2 mean of AP is called “mAP “which is inconsistent with use of “mean” in recall which is called “mean”.

[4.1] “our method is 3% higher” Consider using “is more accurate” instead.

[Throughout the paper] Certain abbreviations are not explained when used for the first time. (RGB, HSV, HOG, LSS, CNN, FPS). YOLOv2 is used without providing information on what it represents.

Author Response

Please refer to "reply letter-reviewer2. Docx" for details.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have answered my all concerns and I recommend this paper for publication.

Author Response

Thank you very much for the reviewer’s careful guidance.

Reviewer 2 Report

The authors have considerably revised their manuscript. In light of the updated comparisons listed in Table 2, the authors should now elaborate on the discussion of results at the end of Section 4. The discussion should revolve around the comparison of mAP, mR and FPS for all baseline methods against their own. The advantages or disadvantages/limitations of the proposed against the state-of-the-art should be clearly documented. Moreover, the Conclusions in Section 5 should be updated accordingly.

Author Response

Please refer to “reply letter-reviewer2. Docx” for details.

Author Response File: Author Response.pdf

Round 3

Reviewer 2 Report

The authors have addressed comments.