Spatial–Temporal Attention Two-Stream Convolution Neural Network for Smoke Region Detection
Round 1
Reviewer 1 Report
Image recognition techniques for smoke and fire detection are the future of fire protection. When refined, this kind of technology will assure reliant and fast alarming with limitation of false alarms.
While already used detection has reliability in closed compartments we still are looking for best solution outside eg. wildfires.
Proposed in presented paper solution is high technological solution and states next step to better understanding of fire recognition problem.
I believe that this kind of solution shouldn’t be described as auxiliary (line 37) – but in near future – the main one.
The article is well written and clear for people who know principles in cybernetics. But might be confusing for fire researchers.
My remarks to the article:
41 – please reconsider change of phrase “handicraft features”
42 – some description of HIS color space is needed
45 – some description of YUV color space is needed
50 – please reconsider change of word “excavation” (maybe “recognition”?)
250 – please clarify what pytorch is? Is this a PyTorch software?
341 – please reconsider change of word “mine”
342 – please add dot at the end of sentence
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper studies a new type of smoke detection model that indeed improves the accuracy of smoke recognition compared to the previous model is of great significance to the detection and prevention of the initial fire. However, there are still some questions that are not clear. Detailed comments are as follows.
- As mentioned, ‘most of the researches focused on smoke image recognition or smoke target location’. Why do we care about the smoke region detection? In other words, do we really need very high accuracy to recognize the boundary of the smoke? What do we need that for?
- In line 141 ‘We treat each pixel of the template frame feature as a template to do correlation calculation with the current frame’, can you elaborate on the method of calculating the correlation between the template frame and the current frame?
- The four datasets in the article are used for training and testing respectively. What criteria are they divided according to?
- Can you explain in detail from the model level why your algorithm is better than the other three algorithms (FCN, Deeplabv3 + and RANet)?
- In line 283 ‘we use the spatial-temporal attention to shield possible noise’, The specific implementation principle may need to be completed.
- Two model evaluation methods (IoU, F1−score) are used in this paper. Why are two methods used to evaluate models?
- Please double check the whole paper to avoid small mistakes or typos, e.g., Line 16 ‘smoke colors of smoke’.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The author states that smoke region detection can help to estimate the fire spread trend and rate. I do not fully agree to be honest. It is true that the smoke trajectory is a clear sign of the wind direction. But it doesn’t need very high resolution to recognize the smoke moving direction, no? The other possible application for estimating the fire spread rate is not reasonable/acceptable in my opinion. The authors should not exaggerate their works and need to think carefully about the contribution of this article.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
