Automatic Detection of Construction Workers’ Helmet Wear Based on Lightweight Deep Learning
Round 1
Reviewer 1 Report
This paper is based on the GhostNet as the light-weighted pretrained backbone for worker's
helmet wearing detection from viusal images. Specifically, a multi-scale segmentation and
feature fusion network (MSFFN) is devised to extract informative features, together with
the feature fusion network and a tailored attention mechanism for the final object position
and category prediction module, which specifically addresses the presence of different
scales of the objects.
There are some specific suggestions to improve the paper:
i) enhancing related work and citations:
the paper in the experiment part mentioned the used dataset: SafetyHelmetWearing-Dataset
(SHWD) and Safety Helmet detection with Extended Labels 5K images (SHEL5K). Yet I
did not find any reference to these datasets which may not be easy for readers to identify
the data sources, be it a public benchmark or not.
In the experiments, the compared methods can be more diverse and some recent methods
especially for small object detection can be considered (open sourced):
1) SCRDet++: Detecting Small, Cluttered and Rotated Objects via Instance-Level Feature
Denoising and Rotation Loss Smoothing, IEEE Transactions on Pattern Analysis and
Machine Intelligence (TPAMI), 10.1109/TPAMI.2022.3166956, 2022
Of course I could understand that it may be nontrivial to comapre these methods in your
setting. Therefore you could at least discuss these recent and representative works in the
related work part (the ICCV19 conference version for SCRDet has 400+ google citations
which cannot be ignored by this work as both handle with small object detection):
2) SCRDet: Towards More Robust Detection for Small, Cluttered and Rotated Objects,
International Conference on Computer Vision (ICCV), 2019.
Specifically, the idea of devising tailored attention is also studied in previous work and thus
you could better discuss them in related work.
ii) For future work, I would suggest the authors deep dive into the recent works on small
object detection e.g. the works mentioned above, and in fact, many of the works provide
plug-in e.g. new loss, new module that can be combined with the presented work.
iii) For the experiments part, the authors could provide the detailed information of the
combined dataset to make the paper more reproducible.
Minor suggestions:
2
1) the source code of this paper is encouraged to be made publicly available 2) it would be
better to discuss any limitation of this work e.g. some failure cases 3) also you should clarify
if your method is single-stage (I think so which is just like the above R3Det (AAAI21))
ii) For future work, I would suggest the authors deep dive into the recent works on small object detection e.g. the works mentioned above, and in fact, many of the works provide plug-in e.g. new loss, new module that can be combined with the presented work.
iii) For the experiments part, the authors could provide the detailed information of the combined dataset to make the paper more reproducible.
Minor suggestions:
1) the source code of this paper is encouraged to be made publicly available 2) it would be better to discuss any limitation of this work e.g. some failure cases 3) also you should clarify if your method is single-stage (I think so which is just like the above R3Det (AAAI21))
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors proposed a novel method for a specific task of detection (and localization) of objects in an image exploiting a deep learning-based technique. The method is exploiting a GhostNet.
The idea is well presented, no issues in the structure of the paper or in the exposition of the content (images, English, tables). Also, the architecture and the experiments respect the quality of a relevant scientific study.
However, some questions should find an answer:
---
Why did you propose a lightweight architecture and then train the system on a mid-high-range computer (powered by an RTX 3050)? Usually, lightweight algorithms are specifically designed for their usage on embedded devices or limited-resources ones, as explained in
Danilo Avola, Luigi Cinque, Alessio Fagioli, Gian Luca Foresti, Marco Raoul Marini, Alessio Mecca, Daniele Pannone:
Medicinal Boxes Recognition on a Deep Transfer Learning Augmented Reality Mobile Application. CoRR abs/2203.14031 (2022)
Thus, which is your aim in this sense? We suggest the authors to underline this fact in the manuscript.
---
In the experimental section, self-evaluation of the effectiveness is performed. However, a comparison with other similar systems in literature should be made too, e.g., Hayat, A.; Morgado-Dias, F. Deep Learning-Based Automatic Safety Helmet Detection System for Construction Safety. Appl. Sci. 2022, 12, 8268. https://doi.org/10.3390/app12168268 or other examples. We suggest the authors to insert a sub-section in which they compare their results with the ones collected by other research groups.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
