Learning Point Processes and Convolutional Neural Networks for Object Detection in Satellite Images
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper presents a novel small target detection framework that combines Convolutional Neural Network pattern extraction with the Point Process method, and introduces Energy Based Model. While the proposed framework has demonstrated promising results in vehicle detection from satellite images, there are several points that require improvement:
1. Lack of conclusion part.
2. This paper contains many formulas and parameters. Of course, the derivation is very rigorous and clear. If possible, can the formula be reduced? This is just a suggestion.
3. In line 51, a citation appears as [?], which needs to be corrected with the appropriate reference.
4. On line 165, the first PP abbreviation should give the full name.
5. Figure 4 should include a mention of the target 'u' for clarity and completeness.
6. In Figure 6, it's recommended to label 'dc' to aid readers in understanding the difference between a cell and a set.
7. In Section 3.2, the paper should mention the parameter settings for all models used in the experiments.
8. In Figure 9, it is suggested that the third sample is to add noise to the first or second row of samples.
9. For the ADS dataset, both qualitative and quantitative comparison results of different models should be included.
10. Section 3.3.1 should include comparisons of computational complexity and the number of parameters with other existing models for a comprehensive evaluation.
11. The discussion section should include an analysis of the shortcomings of the proposed model.
12. Consider discussing how different datasets and vehicles of varying sizes may affect parameter adjustments in the proposed framework.
Author Response
We wish to thank Reviewer 1 for their suggestions and comments, which have helped us to improve the quality of our manuscript. We have carefully revised the paper according to these suggestions and addressed all comments. The modifications are detailed in the attached file.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe manuscript is investigating the use of Point Process (PP) models for object detection with priors. The authors propose using Markov Marked PP model on the CNN features (data term) in an attempt to improve object detection in remote sensing applications.
The conceptualization is well-formulated and written, and the fact that sustainability was considered to conclude an easily parallelizable execution is a plus.
When sampling a new point, the proposed solution to estimate the density of the point at question during the Birth move through the conditional positional energy of the known maps could use more explanation.
It would be useful to have axes labels in Figure 1 and how i and j represent coordinates of the points to model (e.g. objects’ centroids).
The figures of section 2 could be explained more, or at least, associated more with the discussion in that section.
In section 3.1, how many configurations the model suggested in the comparative experiment on that sample image? i.e. how many positive/negative objects were used to compute the precision-recall curves and the average AP? It would be interesting to see how these differ with the cell location on the image.
I am not a fan of naming models with symbols (the diamond and the star). It is a little confusing and gives the false impression of a footnote or highlight. Perhaps simple subscripts for minimal (m) and whole (w) or something similar makes the names more reliable.
It would be better if Table 2 only showed the table, while the figures on the right are moved to a separate figure. It is highly unusual to mix representation of results like this. Plus, the graphs are too small.
Moreover, the colours used to differentiate the models are not very accessible (consider the colour-blind and sight-impaired readers). Perhaps if the figures are enlarged and the curves are directly annotated, that would make it easier to observe the differences.
Is Figure 11c the result of clustering the objects based on score?
The appendices are very useful to complement the foundation of the proposed model.
The "discussion" completely falls short of analysing the work and results. It reads more like a long summary.
The justification given for this work is to improve detection of satellite images with their flaws. The proposed approach is theoretical sound but was not verified. A deeper analysis is needed. A closer look at the FP and FN. The manuscript did not address occlusion or dense locations. Also, any limitations should be highlighted.
Ideally, there should be a conclusions section.
Comments on the Quality of English LanguageA round of proof reading is strongly recommended. Punctuations (particularly commas) are heavily missing, and it is causing confusion in some sentences.
Some spotted edits:
Line 51: missing citation number.
Line 675: “due ^ industrial property restrictions.” – missing a preposition “to”?
Figure 2: the subplots are captioned numerically while the figure caption refers to alphabetic numbering. Please unify.
Line 152: the equation exceeds the page margin.
Line 154: “with V(.) the energy” -> “with V(.) being the energy” or “where V(.) is the energy”.
Line 159: “referred as observation” -> “referred to as observations”.
Line 193: “with [.] the corresponding” -> “with [.] being the” or “where [.] is the”.
Line 357: should theta be boldface (vector of parameters)?
Line 373: subscript of L is lowercase.
Line 437: “energy model ^ minimal energy” – missing a preposition “with”?
Line 457: “require few false positive” -> “require low false positive”.
Line 458: “require less missed detection” -> “require fewer miss detections”.
Please revise the last line in the caption of Figure 8. It is confusing.
Please revise grammar and sentence structure is the caption of Table 2.
Table 1 caption is missing a full stop. Also, it should be above the table.
Line 541: broken link.
Author Response
We wish to thank Reviewer 2 for their suggestions and comments, which have helped us to improve the quality of our manuscript. We have carefully revised the paper according to these suggestions and addressed all comments. The modifications are detailed in the attached file.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe author has revised all comments and I agree to accept the current version.
