Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Advantages:

1.The research content is of great significance for the sustainable development of agriculture.

2.The format layout is beautiful and clear.

3.The language expression is clear.

 

However, there are still several issues that need to be addressed.

Comments:

1.Insufficient references

The number of references in the article is insufficient, and the focus is not clear enough. Especially when introducing the neural implicit surface reconstruction method, only NeRF is mentioned, while other pioneering related algorithms are not introduced,such as NeuS,Volosdf and 3DGS. The related work section also lacks citations related to multiview-images based 3D point or mesh Reconstruction of crop/plants.

2.Insufficient datasets and comparative experiments

The qualitative experiments in the article are only conducted on single pepper dataset which seems to only contain one pepper sample. Why not an average evaluation on a group of peppers?  furthermore，why not qualitative experiments on other fruit datasets？

Also，in Table 3, authors didn't make comparison with some SOTA implicit surface reconstruction methods (such as NeuS,Neuralangelo etc.) and 3DGS methods(such as GoF).

3.No details about data acquisition.

the collection process of the pepper dataset is not described in detail. For example, it is not clear how many pictures are taken for each plant in the collected Pepper dataset, 

- "At nine predetermined positions along this path" is curtly introduced in line 265.

- "The study recorded the exact position and corresponding 3D point cloud at each sampling location. The 3D point cloud data was then registered and fused" (line 288), how to do?  

- How to calculate the poses and intrinsic parameters of motion camera？ 

- How many captured images are able to meet the demand of 3D reconstruction？

- How to deal with the occlusion issues in data collection？

....

The authors are expected to clarify these details.

 

4.unclear description of research results

In the results, some 3D reconstructed meshes have colors and others do not,why?  It is necessary to explain how the colored meshes are achieved. In addition, other charts (such as Fig.8, Fig.9,) lack clear explanations of specific contents, and some chart orders are incorrect, which affects the readability of the article.

 

In Table 5, are the phenotypic measurements (height and width) represented which organic of the pepper plant？(obviously not pepper fruit) 

Is it measured by manual?  if so, 1) which software is used to extract the height and width  2) the 3D semantic segmentation method described in Fig.5 seems unnecessary, because the fruit can be segmented or measured by manual?  

5.Unclear source of point clouds in Chamfer Distance Evaluation

The source of the point clouds used in the Chamfer distance evaluation needs to be further clarified. Especially if the GS-NIR model uses scanned point cloud data during training, it may create an unfair advantage during evaluation.

In addition，it seems unnecessary to conduct quantitative measurements on the reconstructed point clouds in Section 4.4, and there are also problems in the experimental design in Section 4.5. In the experiment, the comparison results with and without scale transformation and point cloud segmentation should be described in detail. 

 

Author Response

Dear Reviewers:

 

Thank you for your letter and for the reviewers’ comments concerning our manuscript entitled “Exploring Accurate Phenotypic Reconstruction of Fruits by Neural

Implicit Surface” (Manuscript ID: agriculture-3344731). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our research. We have studied comments carefully and have made corrections which we hope to meet with approval.

 

Comments 1:

Insufficient references. The number of references in the article is insufficient, and the focus is not clear enough. Especially when introducing the neural implicit surface reconstruction method, only NeRF is mentioned, while other pioneering related algorithms are not introduced,

such as NeuS, Volosdf and 3DGS. The related work section also lacks citations related to Multiview-images based 3D point or mesh Reconstruction of crop/plants.

 

Response 1: We sincerely appreciate your valuable comments. We have carefully checked the literature and added more references on neural implicit surface reconstruction methods in the introduction section of the revised manuscript. These include NeuS, VolSDF and 3DGS.

--------- line (110-111,115-121)

3DGS is a method based on a learning-based Gaussian kernel that yields high-quality renderings but poor quality surface reconstruction.

NeuS and VolSDF proposed the Signed Distance Function (SDF) to improve the quality of surface reconstruction by imposing model constraints.

Instant-NSR and NeuS2 improved the reconstruction speed without degrading the reconstruction quality by combining the advantages of hash coding and SDF.

The efficient computation of numerical gradient for hash grid proposed by Neuralangelo solves the problem of non-smoothness of the learned surface in the hybrid method.

Papers related to multi-view-based plant reconstruction have also been added.

--------- line (96-100)

Jay et al. used a digital camera to capture continuous images and incorporated a SfM approach to reconstruct field crop models.

Zermas et al. extracted the geometry of maize stalks and leaves and extracted phenotypic features through a combination of drone and camera and 3D reconstruction using SfM.

And 3D reconstruction of apples and oranges were reconstructed by a stereo vision system.

 

Comments 2:

Question 1: Insufficient datasets and comparative experiments. The qualitative experiments in the article are only conducted on single pepper dataset which seems to only contain one pepper sample.

Why not an average evaluation on a group of peppers? 

Question 2: Why not qualitative experiments on other fruit datasets？Also，in Table 3, authors didn't make comparison with some SOTA implicit surface reconstruction methods (such as NeuS, Neuralangelo etc.) and 3DGS methods(such as GoF).

 

Response 2:

Question 1 answer: Thank you for your question. Because our data was collected during the growing season of the peppers, the appearance of the peppers remained almost uniform. Also, in this study we only measured the length and width of the pepper, so our data focuses more on the differences in the length and width size of the pepper. We collected 30 sets of pepper data samples for evaluation.

--------- line (304)

Each of the two sets of data was acquired for 30 scenes.

 

Question 2 answer: Thank you for your comments. We considered the qualitative aspects of other fruits. However, in our work in this study, we only focused on peppers for a more detailed analysis. We will consider expanding the testing of qualitative studies in our future work. In the meantime, we tested the surface reconstruction methods of Neus and Neuralangelo in our previous experiments, and they performed very poorly in the pepper scene, so we did not include a comparison. The model we chose to compare (Instant-NSR) was the best performing surface reconstruction model tested in previous studies [1]. For 3DGS, which is a model that focuses on rendering effects, the extraction of surfaces is also very poor.

 

Comments 3: No details about data acquisition.

Question 1 - The collection process of the pepper dataset is not described in detail. For example, it is not clear how many pictures are taken for each plant in the collected Pepper dataset.

Question 2 - "At nine predetermined positions along this path" is curtly introduced in line 265.

- "The study recorded the exact position and corresponding 3D point cloud at each sampling location. The 3D point cloud data was then registered and fused" (line 288), how to do? 

Question 3 - How to calculate the poses and intrinsic parameters of motion camera？

Question 4 - How many captured images are able to meet the demand of 3D reconstruction？

Question 5 - How to deal with the occlusion issues in data collection？

 

 

Response 3:

Thank you for your question. We will describe the experimental data collection section in more detail in the revised manuscript.

Question 1 answer: When we collect image data of peppers, each group of peppers is collected with a high-speed camera and 100 photos are taken for reconstruction.

Question 2 answer: Regarding the point cloud data collection of peppers. This study only uses a structured light scanner to acquire 3D point clouds for result comparison. We does not involve the study of point cloud alignment related methods, it is not described in detail in the manuscript.

Question 3 answer: We use Colmap [2] for the estimation of motion camera pose and internal reference.

Question 4 answer: We use 100 high-speed camera images to meet the reconstruction requirements.

Question 5 answer: There is no serious occlusion problem in our acquired data. The occlusion of fruits by some leaves can be supplemented by other viewpoints, which can also achieve high quality reconstruction results.

 

We added the relevant description in

--------- line (261-266)

The first approach to image data acquisition utilizes a GoPro Hero 11, an action camera with a high frame rate and high resolution, in order to facilitate the acquisition of data in a lightweight and speedy.

This camera is capable of capturing 4K resolution images at a frame rate of 120Hz, ensuring both quality and continuity in the data collection process. For each data set 100 images are extracted. Camera pose estimation as well as internal parameter computation is done using Colmap.

 

Comments 4: Unclear description of research results.

Question 1: In the results, some 3D reconstructed meshes have colors and others do not, why?  It is necessary to explain how the colored meshes are achieved. In addition, other charts (such as Fig.8, Fig.9,) lack clear explanations of specific contents, and some chart orders are incorrect, which affects the readability of the article.

Question 2: In Table 5, are the phenotypic measurements (height and width) represented which organic of the pepper plant？(obviously not pepper fruit)

Question 3: Is it measured by manual?  if so, 1) which software is used to extract the height and width  2) the 3D semantic segmentation method described in Fig.5 seems unnecessary, because the fruit can be segmented or measured by manual? 

 

Response 4:

Question 1 answer: Regarding the mesh colors in Figures 7, 8 and 9. One type of mesh is the mesh with the colors removed, to better demonstrate the quality of the reconstruction of the pepper surface. The other colored mesh is the mesh with normal direction, which shows the accuracy of the reconstructed mesh. Both meshes are extracted from the model reconstruction results. Our results are all meshes with the real colors of the peppers. We have revised the figure annotations and rearranged the image order in the revised version.

Question 2 answer: We were sorry for our careless mistakes. Thank you for your reminder. The unit of measurement in Table 5 is wrongly labelled. The correct unit should be millimeter. The data was measured from real pepper fruits. We have corrected this in the revised version.

Question 3 answer: We used CloudCompare to extract the height and width of pepper in our experiment. We will provide additional explanation in the revised manuscript. Our primary goal with the 3D semantic segmentation approach is to eliminate the interference of branches and leaves on the measurement results. It allows for an automated phenotypic measurement platform.

--------- line (396-398)

Subsequently, to obtain more accurate phenotypic data and to eliminate the interference of branches and leaves on the measurements, the 3D point cloud was segmented using Fuse-PointNet++ and extracted for the single pepper plant to be tested.

 

 

Comments 5: Unclear source of point clouds in Chamfer Distance Evaluation.

Question 1: The source of the point clouds used in the Chamfer distance evaluation needs to be further clarified. Especially if the GS-NIR model uses scanned point cloud data during training, it may create an unfair advantage during evaluation.

In addition，it seems unnecessary to conduct quantitative measurements on the reconstructed point clouds in Section 4.4.

Question 2: there are also problems with the experimental design in Section 4.5. In the experiment, the comparison results with and without scale transformation and point cloud segmentation should be described in detail.

 

Response 5:

Question 1 answer: In the evaluation of chamfer distance, we use the point cloud acquired by a structured light scanner as a reference value. Since our goal is to use a lightweight high-speed camera instead of an expensive scanner, we chose to evaluate the point cloud with the scanner.

--------- line (372-375)

In this section, a point cloud of peppercorns was reconstructed using a high-precision 3D scanner, and this point cloud was compared with GS-NIR, Instant-NGP, and Instant-NSR as ground truth.

 

Question 2 answer: In the experiment of 4.5, the results before scale transformation were not compared. This is because with the current NeRF-based reconstruction methods, the scales are redefined by Colmap when calculating the camera pose and internal parameters. There is no reference value for real scenes because the scale is recalculated in a fixed space no matter how big the scene is. Therefore, we need to restore the scale of the reconstructed model to continue the measurements.

--------- line (407-409)

The data in table 5 compare the three models and the 3D scanner with real measurements after scale restoration and point cloud segmentation.

 

 

References:

[1] Hu, K.; Ying, W.; Pan, Y.; Kang, H.; Chen, C. “High-fidelity 3D reconstruction of plants using Neural Radiance Fields.” Computers and Electronics in Agriculture 220 (2024): 108848.

[2] Schonberger, Johannes L., and Jan-Michael Frahm. "Structure-from-motion revisited." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.

 

Reviewer 2 Report

Comments and Suggestions for Authors

The following review provides a critical analysis of the article, which explores the use of neural implicit surfaces for accurate phenotypic reconstruction in agriculture, to highlight its strengths, address potential areas for improvement, and suggest enhancements to strengthen its contribution to the field of precision agriculture.

1 - The abstract could explicitly highlight the broader impact of the study on precision agriculture, particularly its scalability for diverse crops. Avoid using technical terms like "NeRF" without brief contextualization in the abstract for a multidisciplinary audience.

2 - The introduction effectively sets the context by highlighting the importance of accurate plant phenotyping in precision agriculture and the limitations of traditional methods like manual measurement and conventional 3D scanners. It ties the study to emerging technologies like Neural Radiance Fields (NeRF) and explains their potential in overcoming existing challenges in phenotyping. The text clearly articulates the limitations of current methods (e.g., cost, complexity, limited scalability) and justifies the need for the proposed GS-NIR model. However, while the introduction mentions that GS-NIR offers improved scalability and robustness, it does not explicitly explain how it achieves this or provide examples of its expected impact. The introduction could benefit from a broader discussion of the global challenges in sustainable agriculture and how the study contributes to addressing these issues. Suggestions: (i) Provide a brief overview of the GS-NIR's core innovations in lay terms, such as its ability to handle complex plant geometries or operate in less controlled environments; (ii) Discuss how this research aligns with broader trends in AI and agriculture, such as automation and environmental sustainability.

3 - The methodology section is detailed, providing a clear breakdown of the GS-NIR model's architecture and its enhancements, such as Geometric Depth Smoothing Constraints and Random Dilation Sampling. The inclusion of Adaptive Scene Modules and the combination of Hash Table Encoding (HTE) and Spherical Harmonics Encoding (SHE) demonstrate a well-thought-out approach to addressing specific challenges in agricultural scenarios. But some sections, particularly those involving mathematical formulas, may be difficult for non-specialists to follow. For example, the equations describing depth smoothing constraints could benefit from additional explanatory text or visual aids.

4 - The study presents a comprehensive comparison of GS-NIR with state-of-the-art methods like Reality Capture, Instant-NGP, and Instant-NSR using metrics such as PSNR, SSIM, and Chamfer Distance. These studies effectively demonstrate the impact of individual components (e.g., depth smoothing, sampling strategies) on reconstruction quality. Also, the use of figures to showcase reconstruction results and the effects of various optimizations provides clear evidence of the model's performance. But the discussion focuses heavily on technical metrics but provides limited insight into how the improvements translate into practical agricultural benefits. So it is important to discuss the practical implications of reduced errors and improved reconstruction quality in terms of cost, efficiency, or ease of adoption in real-world agricultural scenarios. Also, to expand on the versatility of GS-NIR by presenting results for other crops or under different environmental conditions, as hinted at in the experiments.

5 - Figures are informative and well-labeled, particularly those showcasing 3D reconstructions. Enhance clarity by bolding key results in tables (e.g., Table 3) and ensuring uniform formatting. Consider annotating figures to emphasize specific improvements (e.g., fewer holes or better mesh quality).

6 - The conclusion effectively summarizes the study's contributions and outlines future directions. Provide specific recommendations for future research, such as extending GS-NIR to other crop types or improving robustness in outdoor environments. Highlight the scalability of the proposed method in terms of cost-effectiveness or ease of deployment in diverse agricultural settings

7 - Proofread for minor grammatical issues and ensure consistent use of technical terms throughout the paper.

Author Response

Dear Reviewers:

 

Thank you for your letter and for the reviewers’ comments concerning our manuscript entitled “Exploring Accurate Phenotypic Reconstruction of Fruits by Neural

Implicit Surface” (Manuscript ID: agriculture-3344731). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our research. We have studied comments carefully and have made corrections which we hope to meet with approval.

 

Comments 1:

The abstract could explicitly highlight the broader impact of the study on precision agriculture, particularly its scalability for diverse crops. Avoid using technical terms like "NeRF" without brief contextualization in the abstract for a multidisciplinary audience.

 

Response 1:

Thank you for your comments. Therefore, we have made relevant additions and revisions.

---------(line10-13)

The mean distance error between the scanner-based method and the NeRF(Neural Radiance Fields)-based method is 0.811mm.

This study shows that the learning-based NeRF method has similar accuracy to the 3D scanning-based method, but with greater scalability and faster deployment capabilities.

We describe the technical background of NeRF in more detail in our related work.

--------- line (110-111,115-121)

3DGS is a method based on a learning-based Gaussian kernel that yields high-quality renderings but poor quality surface reconstruction.

NeuS and VolSDF proposed the Signed Distance Function (SDF) to improve the quality of surface reconstruction by imposing model constraints.

Instant-NSR and NeuS2 improved the reconstruction speed without degrading the reconstruction quality by combining the advantages of hash coding and SDF.

The efficient computation of numerical gradient for hash grid proposed by Neuralangelo solves the problem of non-smoothness of the learned surface in the hybrid method.

 

Comments 2:

The introduction effectively sets the context by highlighting the importance of accurate plant phenotyping in precision agriculture and the limitations of traditional methods like manual measurement and conventional 3D scanners. It ties the study to emerging technologies like Neural Radiance Fields (NeRF) and explains their potential in overcoming existing challenges in phenotyping. The text clearly articulates the limitations of current methods (e.g., cost, complexity, limited scalability) and justifies the need for the proposed GS-NIR model. However, while the introduction mentions that GS-NIR offers improved scalability and robustness, it does not explicitly explain how it achieves this or provides examples of its expected impact. The introduction could benefit from a broader discussion of the global challenges in sustainable agriculture and how the study contributes to addressing these issues. Suggestions: (i) Provide a brief overview of the GS-NIR's core innovations in lay terms, such as its ability to handle complex plant geometries or operate in less controlled environments; (ii) Discuss how this research aligns with broader trends in AI and agriculture, such as automation and environmental sustainability.

 

Response 2:

Thank you for the valuable feedback. We add a description of artificial intelligence and precision agriculture as relevant. and summarize our three core approaches.

---------line (16-17)

Artificial intelligence has facilitated the rapid development of precision agriculture, which is significantly more efficient and productive than traditional production methods.

---------line (58-64)

Introduce a novel implicit surface network, Geometry-Smooth Neural Implicit suRface (GS-NIR), focusing on improving the quality of geometry reconstruction, showing an enhanced performance in fruit phenotyping.

Deploy a sparse neural network for instance segmentation on reconstruction results and process phenotypic data.

Implement a robotic data acquisition and phenotyping process based on the above two innovations.

 

Comments 3:

The methodology section is detailed, providing a clear breakdown of the GS-NIR model's architecture and its enhancements, such as Geometric Depth Smoothing Constraints and Random Dilation Sampling. The inclusion of Adaptive Scene Modules and the combination of Hash Table Encoding (HTE) and Spherical Harmonics Encoding (SHE) demonstrate a well-thought-out approach to addressing specific challenges in agricultural scenarios. But some sections, particularly those involving mathematical formulas, may be difficult for non-specialists to follow. For example, the equations describing depth smoothing constraints could benefit from additional explanatory text or visual aids.

 

Response 3:

Our key modules, Geometric Depth Smoothing Constraints and Random Dilation Sampling, are demonstrated in Figures 3 and 4. It is possible that the diagram descriptions are not clear enough, which makes your understanding difficult. We will re-do the diagram descriptions to give the reader a clearer explanation.

 

 

Comments 4:

The study presents a comprehensive comparison of GS-NIR with state-of-the-art methods like Reality Capture, Instant-NGP, and Instant-NSR using metrics such as PSNR, SSIM, and Chamfer Distance. These studies effectively demonstrate the impact of individual components (e.g., depth smoothing, sampling strategies) on reconstruction quality. Also, the use of figures to showcase reconstruction results and the effects of various optimizations provides clear evidence of the model's performance. But the discussion focuses heavily on technical metrics but provides limited insight into how the improvements translate into practical agricultural benefits. So it is important to discuss the practical implications of reduced errors and improved reconstruction quality in terms of cost, efficiency, or ease of adoption in real-world agricultural scenarios. Also, to expand on the versatility of GS-NIR by presenting results for other crops or under different environmental conditions, as hinted at in the experiments.

 

 

Response 4:

Thank you for your suggestions. The content of the study results and the benefits of the actual agricultural scenarios have been reorganized.

--------- line (417-420)

Phenotyping researchers no longer need expensive and inconveniently portable scanners. Instead, they can quickly collect phenotypic data in a short period of time by carrying only lightweight camera equipment.

 

Comments 5:

Figures are informative and well-labeled, particularly those showcasing 3D reconstructions. Enhance clarity by bolding key results in tables (e.g., Table 3) and ensuring uniform formatting. Consider annotating figures to emphasize specific improvements (e.g., fewer holes or better mesh quality).

 

Response 5:

We will standardize the formatting of the tables. Figure annotations will also be improved to ensure that readers can clearly understand the results.

 

Comments 6:

The conclusion effectively summarizes the study's contributions and outlines future directions. Provide specific recommendations for future research, such as extending GS-NIR to other crop types or improving robustness in outdoor environments. Highlight the scalability of the proposed method in terms of cost-effectiveness or ease of deployment in diverse agricultural settings

 

Response 6:

Regarding extended studies on other crop types, we will continue to test and refine them in future work. We will add a description of the cost-effectiveness of the methodology of this study.

 

Comments 7:

Proofread for minor grammatical issues and ensure consistent use of technical terms throughout the paper.

 

Response 7:

Thanks for your careful checks. We perform detailed grammar proofreading to ensure that terminology is consistent.