Review Reports

Round 1

Reviewer 1 Report

This work proposes a driver fatigue detection system, based on Residual Channel Attention Network (RCAN) and Head Pose Estimation. The authors employ Retinaface for face location in order to detect five face landmarks. Then RCAN is adopted to classify the state of eyes and mouth, including a channel attention module. In particular, a PERCLOS and POM is adopted to judge whether the driver is in the state of drowsiness. The authors also compare the Euler angle output by EPNP method with the Over-angle of head pose to judge whether it is in the state of over deflection of head.

Issues to fix to improve the quality of the paper:

1. This article mainly contains three aspects: facial state recognition, head pose estimation and fatigue assessment. The Section 1. Introduction does
not provide a sufficient overview of the topics analyzed. The Basics of the
related studies should be reinforced. So, the authors should consider the
chance to check if it can be inserted a section “Related works”.

2. In literature, there are a myriad of approaches to estimate the head pose
(See Point 1). What are the reasons behind the decision to adopt the
EPNP method for HPE?
See the following recent works (doi):
• 10.1007/s00138-021-01234-1
• 10.1016/j.patrec.2020.10.003
• 10.1109/TPAMI.2020.3046323

3. Although the study appears to be sound, the RCAN and its channel attention module description are unclear and difficult to follow. I suggest
to explain it with less mathematical formalism (the same goes for HPE
module).

4. In the experimental phase, several works related to “Driver Fatigue Detection” adopt the NTHU-DDD Dataset, a driver drowsiness video database collected by NTHU Computer Vision Lab. So, the authors should consider the chance to check if it can be used the above-mentioned benchmark.

5. In order to prove the superiority of RCAN in driving fatigue detection, the authors compared RCAN with other classical CNN structures, i.e VGG-16, ResNet50 and Inception. However, the bibliographic references reported are not very recent (2014 and 2016, respectively). In fact, in the last four years, several works are published in this field. Please improve this important aspect!

6. In the final section, authors should analyze the limitations of the proposed system (if they exist). For example, can it be adopted for real world applications? Which direction will future works focus on?

7. The bibliographic references are not sufficient and the technical depth is
not enough. There are several typos in the paper and the formatting of
the manuscript should be better fixed (i.e tables, equations, bibliographic
references). A deep proofread is suggested.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear authors, I ask you to answer a number of questions and comments:
1. For such a significant problem, the list of references is too small and it is necessary to delete all publications older than 10 years, since their relevance is questionable.
2. What to do with drivers with congenital or acquired disabilities. What is the possibility and likelihood of recognition.
3. The annotation needs to be expanded and made more informative.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Authors' response to my concerns was convincing. 

I found 2 typos, so I suggest a proofread.

Row 93: "Paul vlola et al."
Row 600: "3009 open mouth images" should be 2874

Author Response

 We have corrected the above errors and proofread the full article, found and 
corrected several more errors, such as changing the original "a = 0" to "a = 0.5" in Section 4.3, and three Over-angle values errors in section 4.4. The above errors have been corrected. Thank you for your proposal and review.

Author Response File: Author Response.pdf