Multi-Scale Feature Pyramid Network: A Heavily Occluded Pedestrian Detection Network Based on ResNet

Round 1

Reviewer 1 Report

This paper presents a technique for the detection of pedestrian in occluded scenes. To achieve this, the proposed method builds a multi-scale feature pyramid network based on ResNet (MFPN) to improve the detection accuracy. The overall framework looks technically sound and the reported results are promising.

My one of the concerns is the novelty of the paper. It would be good if authors could please clearly highlight how their architecture is different from the existing networks/models?

I am also concerned about the evaluation of the proposed technique only on one dataset, which is small in size. It would be good to see results on other datasets as this will give confidence to the reviewer and readers about the robustness and accuracy of the proposed technique. 

The quality of presentation can be improved by moving Figures and their captions on the same page. 

I encourage the authors to address these comments and resubmit this article.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

This work presents a neural network architecture based on ResNet, aiming to address heavily occluded pedestrian detection. The proposed multi-scale feature pyramid network (MFPN) includes two modules, double feature pyramid network (FPN) and repulsion loss of minimum (RLM). Experimental comparisons with state-of-the-art methods are performed on the publicly available CrowdHuman dataset.

Although the technical contributions of this work are rather incremental, the manuscript addresses an important problem and could be of interest to the readers of Sensors. Still, the current version suffers in terms of clarity, whereas the presentation of experimental results is not sufficient.

The actual function of the main elements introduced in the proposed architecture is not well explained in the text. E.g. the authors state “the separable convolution of the original BiFPN is replaced by general convolution, which can better extract and fuse the features” (lines 199-201). Why is this so?

Section 3.3: The novel element introduced in the loss function, when compared to [40], is not clear.

Figure 4: it is not clear at all why one feature map is better than another, as stated in the caption.

The results presented in Table 3 are rather marginal and there is no statistical significance analysis.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have revised the paper based on the previous comments and made the required changes in the paper. They have also evaluated their approach on a new dataset and compared with state of the art methods.

The paper can be accepted, however, a quick proof reading will help in further refining the final manuscript.

Reviewer 2 Report

The manuscript has been substantially revised. I am glad to recommend publication.