Object Detection Based on Center Point Proposals

Round 1

Reviewer 1 Report

The authors have presented an approach to increase the accuracy of object detection. Their proposal proposal is based on the objects center point detection instead of anchor-based detectors.

This interesting proposal is based on the combination of two anchor-free detector methods, FCOS and OAP that include some modification to solve/reduce the disadvantages of these implemented techniques, such as:

• FOCS may produce a large number of negative samples, a common error when small targets are in close proximity to each other, i.e. in the same cell, only one will be detected.
• OaP is based on the object's center point, however if the targets are of the same scale and are very close to each other, their Ground Truth center point will overlap during the down-sampling, which leads to a misleading training because it can only train two objects as one.

In the proposed algorithm combines the FOCS that implements the Darknet-53 structure as backbone. The object detection algorithm is based on center point proposals to fill the cells. Then, a differentiation module (DM) reduces the missed detections on the heatmap. Finally, a new loss function, Centerness-H (CH) is applied to the center point predicted by the heatmap, instead of to the center of the cells.

The author have presented a well organized and explained paper. They have properly addressed the analysis of the state-of-the-art methods, as well as the experimentation section. They have covered the most relevant experiments to compare the CPO with other methods. For instance, the implementation of different backbones and methods, and the combination of the proposed modules DM and CH.

The results shown that the proposed method slightly outperform the other detectors. FCOS, CPO-R and CPO (Darknet-53) have almost the same performance. Then, perhaps the relevant issue would be how good is the performance of the proposed method when processing challenging images. It would be very interesting for authors to include more complex images/scenes in the experimental test results, perhaps images similar to those presented in section 1, Figure 3.

Author Response

Dear reviewer:

 

I am very grateful to your comments for the manuscript. All these advices have been considered and responded carefully. According with your advices, we amended the relevant part in our manuscript. We hope that the quality of this manuscript can be improved by our modification. Point-by-point responses to the comments are listed below.

 

Point 1: Then, perhaps the relevant issue would be how good is the performance of the proposed method when processing challenging images. It would be very interesting for authors to include more complex images/scenes in the experimental test results, perhaps images similar to those presented in section 1, Figure 3.


 

Response 1: Thanks for your suggestions. Your comment is very helpful to us. According to your suggestion, we have added more complex images/scenes in the experimental test results. those are presented in section 4, Figure 8. Line 428-435 in our revised manuscript.

 

 

Thank you again for your positive comments and valuable suggests to improve the quality of our manuscript.

 

 

Best regards.

Yours sincerely,

Hao Chen

Corresponding author:

Name: Hao Chen

E-mail: [email protected]

Reviewer 2 Report

An object detection based on center point proposals is implemented. The proposed scheme may further reduce the number of useless anchor boxes, improve the quality of the anchor boxes, and balance the proportion of positive and negative samples. The presentation is clear with fine simulations. I recommend the acceptance of this paper.

Author Response

Dear reviewer:

 

I am very grateful to your comments for the manuscript. Thank you for good comments and hard work.

 

 

Best regards.

Yours sincerely,

Hao Chen

Corresponding author:

Name: Hao Chen

E-mail: [email protected]

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper provides a good motivation for the research presented within, explaining the target detection problems encountered in several other approaches.

I have some remarks:

• at line 86, The 2 sentences should be combined in only one.
• The ideas on lines 95,97,99 should be bulleted
• Line 99 - the COCO dataset should be referenced.
• Line 125: "me" - has no place there
• Line 126: reference needed for "the step of region proposals is removed by some researchers"
• Line 183: shouldn't "sample" be replaced with "label" (at the beginning of the line)
• Line 279: the reference [7] should be replaced with a reference to the formula (7)
• The caption in Table 1 should contained detailed explanation of the meaning of the columns in the table
• Line 321: the sentence lacks meaning: "Because the topology of ResNeXt-32x8d-101 is more conducive to feature separation and
  extraction of small targets."
• REference [23] lack the numbering.

Author Response

Dear reviewer:

 

I am very grateful to your comments for the manuscript. All these advices have been considered and responded carefully. According with your advices, we amended the relevant part in our manuscript. We hope that the quality of this manuscript can be improved by our modification. Point-by-point responses to the comments are listed below.

 

Point 1: at line 86, The 2 sentences should be combined in only one.


 

Response 1: Thank you for your reminding. We are very sorry for the quality of the language. According to your comments, we have combined “To address the above-mentioned target detection problems, this study proposes an object detection method based on center point proposals by integrating the advantages of FCOS and OaP (As shown in Fig. 2 (b), OaP algorithm could predicts the center point on the heatmap, facilitating positioning and detection of the cell).” at line 100-105.

 

Point 2: The ideas on lines 95,97,99 should be bulleted.

 

Response 2: Thanks for your suggestions. Your comment is very helpful to us. According to your suggestion, the ideas on lines 111,113,115(Original 95,97,99) have been bulleted.

 

Point 3: Line 99 - the COCO dataset should be referenced.

 

Response 3: Thank you very much for your careful review. Line 115(Original 99) - the COCO dataset has been referenced.

 

Point 4: Line 125: "me" - has no place there.

 

Response 4: Thank you for your reminding. We are very sorry for the quality of the language. According to your comments, We modify this sentence to “in most of the early anchor-based detectors, a set of sparse regions of interest (RoIs) are generated and then classified by the network, which is what we called “two-step method” ” at line 143-145.

 

Point 5: Line 126: reference needed for "the step of region proposals is removed by some researchers".

 

Response 5: Thanks for your suggestions. Your comment is very helpful to us. According to your suggestion, we have added relevant references at line 146 (Original 126).

 

Point 6: Line 183: shouldn't "sample" be replaced with "label" (at the beginning of the line).

 

Response 6: Thank you for pointing this out. "sample" has been replaced with "label" at line 205 (Original 183).

 

Point 7: Line 279: the reference [7] should be replaced with a reference to the formula (7).

 

Response 7: Thank you very much for your careful review. The reference [7] has been replaced with a reference to the formula (6) at line 312 (Original 279).

 

Point 8: The caption in Table 1 should contained detailed explanation of the meaning of the columns in the table.

 

Response 8: Thanks for your suggestions. Your comment is very helpful to us. According to your suggestion, we have added more explanations about the columns at line 350-352.

 

Point 9: Line 321: the sentence lacks meaning: "Because the topology of ResNeXt-32x8d-101 is more conducive to feature separation and extraction of small targets.".

 

Response 9: Thank you for pointing this out. We have added relevant references at line 362 (Original 321).

 

Point 10: Reference [23] lack the numbering.

 

Response 10: Thank you very much for your careful review. 23 has been added.

 

 

 

 

 

Thank you again for your positive comments and valuable suggests to improve the quality of our manuscript.

 

 

Best regards.

Yours sincerely,

Hao Chen

Corresponding author:

Name: Hao Chen

E-mail: [email protected]

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The revised paper is very much improved. All my previous concerns were tackled by the authors.