Infrared Target Detection Based on Interval Sampling Weighting and 3D Attention Head in Complex Scenario
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsSUMMARY:
Algorithm for object detection in infrared thermal images. Low resolution, occlusions. Method with three parallel processes, inspired in YOLO.
The "related work" sections seems more a "foundations of this work" section. I cannot see an extensive comparison with other options for this problem.
GENERAL COMMENTS:
Adhere to MDPI general structure for articles: introdcution (including state of the art revision), materials and methods, results and discussion, conclussion scan be a part of discussion or a standalone section).
Authors choose weighting extracted features instead of downsampling, trying to avoid the discarding of important information.
They apply the same philosophy in the detection head.
It seems an interesting experiment but it is possible that intorduces complexity to get the same results if weighting values end by effectively discarding some information.
A subsection explaining how different channels are obtained and, perhaps, graphical scheme depicting high level overall system structure would be desirable.
DETAILS:
Line 235: "(specifically, pooling layer of 1313, 99, and 55)", IS THIS 1313 kernel size or 13x13, 9x9, 5x5...
Please revise caption of figure 3: 11 or 1x1...
Line 302: "annotated resolution 640512 in the dataset", it should be 640x512, isn't it? The same at 310, 323.
Line 332: "The Average Precision (AP) characterizes the average of the model’s predicting precision over different classes,"
This definition seems not coherent with equation 18, please check.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper presents a framework for the challenging task of the infrared target detection. The main question addressed by the research is how to achieve efficient feature representation in the infrared target detection, to improve its accuracy, and to achieve robustness and anti-occlusion.
The topic has high relevance in the field because although the infrared target detection is well studied due to its practical relevance, the background noise interference and target occlusion loss information pose a challenge for object detection on infrared images.
Several innovative techniques (interval sampling weighting, 3D attention head, C2f module) were added in the framework to improve the accuracy of the detection compared with other published methods. It can provide reliable detection even in case of complex scenarios and occlusions.
A thorough evaluation is performed to compare the performance of the proposed system with popular infrared detection methods.
I miss the evaluation of the running time of the proposed method. It is crucial from the aspect of applicability in real-time. I recommend comparing the running time of the method with the running time of other methods.
The conclusions well summarizes the paper. However, the claim "this algorithm overcomes the obstacles posed by thermal infrared images" is too optimistic: As the second paragraph of the Conclusions states, there are still many challenges in infrared target detection.
The paper is carefully written. The mathematical equations are correct.
The references are appropriate.
I have some further comments:
Please define acronyms at their first occurrence.
Please define C2f.
The font size of the labels in Fig. 1 is too small.
Figure 2, caption: C2f2 - Did you mean C2f?
Line 188: First, The -> First, the
Eq (10): IoU -> IOU.
Table 5, "the best data are described in bold": No bold data in column Bicycle. Two data are described in bold in column 1.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
