Review Reports

Round 1

Reviewer 1 Report

The authors presented the problem of annotation workload which is the need of the hour in yield estimation. The topic is interesting, contemporary and fits with the journal's scope. There are some small issues that authors have to solve to enhance the quality of their publication.

Following include the encountered issues:

1. Abstract lacks a clear summary of what the authors concluded with the published literature.

2. Was the proposed method implemented with other crops other than apple?

3. Is there any mathematical proof to find the center point of the fruit?

4. The application of the proposed method to yield estimation can be elaborated.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear authors,

 

Nice work on fruit detection and counting. The idea of using center point labels to (partially) replace full labels can help facilitate the whole labeling process. The writing is clear and easy to follow. Overall, It’s a good article. To further improve the quality of this manuscript, I have the following questions and suggestions.

 

 

1. The point annotation concept has been proposed in crowd counting for some time. E.g., Ma, Z., Wei, X., Hong, X., & Gong, Y. (2019). Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 6142-6151). A brief review of such concept, method, results, and their relationship to fruit detection should be included in the literature review (not necessarily the paper above).
2. One of the main contributions of using point labeling is reducing the labeling time. Point labeling is for sure much faster than pixel-based annotation but the fair competitor to point labeling should be bounding box labeling (since they both provide the location). How faster is point label compared to bounding box labeling?
3. Since Mask-RCNN provides instance segmentation results which are much harder to learn than location-only results. The fair comparison of your method should be a network with a easier task (just detection than instance segmentation, such as YOLO or SSD).
4. Your method requires a Gaussian mixture model to obtain the center points which require the number of fruits to be accurate. How accurate is your network’s prediction of the number of fruits?
5. The Figures in the manuscript should have higher resolution.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors proposed a semantic graphics (point-to-point) based annotation technique for fruit counting and yield prediction. This technique reduces labelling time and human effort involved and is fast. The paper is technically sound, but a few things must be addressed before the manuscript is ready.

The abstract needs to be rewritten, the authors should focus on the highlights of their paper. The first few lines make abstract sound like Introduction.

Introduction/ Related work section should be improved by citing recent works in the field and how the proposed method is improvement upon those works. Some of the most relevant articles include [1].

Figure 2 needs to be replaced,

·         first it’s very low resolution.

·         The image looks like simply copied from the original paper authors should either get proper copyrights to use it or draw it themselves.

·         Overall pipeline of the network is also broken.

·         It should be clearly drawn for readers to understand it better.

Authors should provide implementation details like batch size, learning rate, no. of parameters, inference speed etc. for better reproduction.

Authors should provide a comparison of how their improved loss function increases the network’s performance.

Did the authors use augmentation? As mentioned in [2] augmentation can have huge impact on networks performance. Experiments can be done to show improvement.

The authors can draw PR-ROC curve for both networks to show the improved performance.

Conclusion needs to improve.

[1]. A Deep Learning Based Approach for Strawberry Yield Prediction via Semantic Graphics.

[2]. DIANA: A deep learning-based paprika plant disease and pest phenotyping system with disease severity analysis

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The work addresses a proposal for fruit detection comparing with the Mask-RCNN method. Using a dataset where the Mask-RCNN method was used. A method of annotation of fruits by central point is also discussed to speed up the annotations for training. Such an approach is not new, but I believe that efforts to improve detection as well as speed up annotations for training neural networks are valid for the fruitculture area.

 

In line (55) a quick annotation method is proposed, indicating only the center of the fruit, as a point. In this publisher (MDPI - doi- 10.3390/rs13010054) there is an article with the same proposal. I suggest a comparison to inform the difference in the method used or indication of the same method.

In figure 2 the resolution of the DLA-34 scheme is small, making it difficult to understand.

In section 3 results and discussion the comparison is made exclusively with the Mask-RCNN method (from reference [18]). There is no comparison of other works with similar, worse or better results than yours. I believe that a comparison with literature results could provide future readers of the work with resources to rank its methodology. If this is not the case, then make it clear in the methodology that the comparison will only be with the Mask-RCNN.

Even so, in this section, the discussion of the results requires links with results from other authors.

It could have a sub-section just comparing results with different methods with the same purpose, showing even the types of sample annotations, box type and showing the differential of the method used as a point.

This section is the one that needs improvement to validate the work.

 

Table 2 - The Title of the table is the same as in the Template. Please correct.

 

Section 4 - It's like "discussion". I believe the correct word would be "Conclusion".

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The authors should use high resolution images in the manuscript (Figure 1).

While submitting review upload both files with and without track changes. Right now the file has track changes ON and is in pdf format and I can't make heads or tails of it.

The authors should add with and without augmentation results as mentioned previously [2]. As it shows adaptability and generalization improvement of the network.

From the answer to review form the authors have answered most the concerns.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The work improved after the suggested corrections, I believe that the research results will be clearer to readers.

Good job.

Author Response

Thanks.