Drivable Agricultural Road Region Detection Based on Pixel-Level Segmentation with Contextual Representation Augmentation

Round 1

Reviewer 1 Report

This research paper proposes a high-resolution detection network designed to precisely segment unstructured road areas to facilitate the autonomous navigation of agricultural robots. Unstructured road boundaries often lack clear semantic identification when compared to their surroundings. To address this, the paper utilizes pixel-level contextual representations at various scales to enhance the identification of road objects. The HRNet (High-Resolution Network) serves as the network's backbone, ensuring the preservation of high-resolution features during downsampling to capture fine details of the roads.

To distinguish drivable and non-drivable regions more effectively, the network incorporates a relational context-based OCR (Object-Contextual Representation) module to enhance feature representations. This module allows distinguishing between the features of road regions suitable and not suitable for driving. Finally, the network employs an adaptive threshold-learned DB (Decision Boundary) head to segment road boundaries accurately.

To evaluate the performance of their method, the researchers created an agricultural unstructured road dataset and conducted experiments. The results demonstrate significant success, with a mIoU (mean Intersection over Union) of 97.85%, indicating the overall segmentation accuracy, and a Boundary IoU of 90.88%, which measures the quality of boundary detection. Comparing the outcomes with existing methods, the proposed network exhibits superior segmentation accuracy and boundary quality, making it well-suited for accurately detecting unstructured drivable areas. 

Review:

The paper covers a meaningful and relevant topic ideally suited to the "agriculture" journal. The authors did a great job of presenting their research. It is very well-written and structured. Conciseness is also a relevant characteristic of this paper. The proposal seems innovative, and the results are convincing, supported by a solid experimental setup. I find the paper technically sound. Congrats.

Author Response

The authors sincerely thank Reviewer #1 for the time, effort, and suggestions given to our manuscript.

Reviewer 2 Report

Overall, an interesting topic. However, there are still some points to improve.
In the abstract, the data source (image recognition from camera) and the "streets" should be specified.
It becomes clear only very late in the text that it is not about field paths but about roads and fixed paths in greenhouses. These roads are not as different from normal roads as the title and the introduction suggest. Please insert clear definitions much earlier (title and introduction).
The study should better explain which features are used to distinguish them. If one cleans the text by the AI "talk" little information remains.
In addition, I would have liked to see the same roads used at different sun positions, as this is important for robot navigation in greenhouses.

Author Response

The authors sincerely thank Reviewer #2 for the time, effort, and suggestions given to our manuscript. All the responses have been shown in the .docx file.

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript describes a deep learning network model for realizing high-resolution segmentation of agricultural roads by leveraging contextual representations to augment road objectness. 

My comments:

1) Regarding Line 130-134: If the reviewer has understood it correctly, there are multiple novel designs throughout the manuscript. Please add ablation analysis to study the separate advantages of each component, e.g., HRNET, OCR, DB decision head, the designed loss function, etc.. 

2) The abstract is too long, where the authors do not need to provide too many details (only general desription is fine). Please shorten it to around 15 lines if possible. 

Language is generally fine, but minor editing is required. 

Author Response

The authors sincerely thank Reviewer #3 for the time, effort, and suggestions given to our manuscript. All the responses have been shown in the .docx file.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

The reviewer thanks the authors to make such changes. The manuscript looks better and it is recommended to be accepted as it is. 

The language requires minor editing.