Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper has a well-structured framework, comprehensive content, strong innovation, and high practical value. The following minor suggestions are provided for your reference:

1. It is recommended to enrich the overall system framework in Figure 4 by incorporating the "deep point cloud-based Continuous Frame Dynamic Constraint Method" into the diagram. Additionally, using experimental images as input, along with intermediate and final results, would make the overall system framework clearer.

2. In Figure 14, it is advisable to add text labels in different colors for better clarity.

3. The paper lacks a discussion section, which should include an analysis of the experimental results, an exploration of existing issues, and an outline of future research plans.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors This paper proposes a low-resource semantic mapping framework for public indoor service robots, addressing high computational overhead and semantic redundancy in existing methods. It designs the lightweight MS-YOLO model (MobileNetV4 + SHViT-based C2f_SHViTBlock), develops a continuous frame dynamic constraint method for redundant annotation elimination, constructs two custom indoor datasets (Indoor1/2), and validates the framework via deployment on the Jetson Orin NX platform. Comparative experiments show MS-YOLO outperforms lightweight YOLO variants in segmentation accuracy with fewer parameters/FLOPs. However, the research suffers from critical limitations in innovation, dataset design, experimental rigor, and real-scene robustness, failing to meet publication standards for now.  

 Strengths

1. MS-YOLO is optimized for resource-constrained service robot edge platforms, with ablation experiments validating the independent and synergistic effects of its core modules.
2. The continuous frame dynamic constraint method effectively addresses multi-view redundant labeling, with a complete semantic landmark construction pipeline (point cloud acquisition → constraint verification → clustering → bounding box fitting).
3. The self-constructed Indoor1/2 datasets fill the gap of public datasets for typical indoor service robot objects, with comprehensive data augmentation to enhance model generalization.
4. End-to-end integration of geometric/semantic mapping pipelines is achieved and deployed on a physical robot with Jetson Orin NX, verifying practical implementability beyond simulation.

Weaknesses

1. Lack of substantive innovation: MS-YOLO is a simple patchwork of existing off-the-shelf modules; the continuous frame constraint and semantic mapping pipeline adopt conventional, widely used designs with no original technical breakthroughs.
2. Small scale and narrow category coverage of custom datasets; no annotation quality control (e.g., inter-annotator agreement); no validation on mainstream public indoor datasets (e.g., NYUv2, ScanNet), leading to unconvincing generalization results.
3. Insufficient comparisons with state-of-the-art lightweight segmentation/semantic mapping models; missing core quantitative metrics for semantic mapping (e.g., ATE/RPE, inference FPS, redundancy elimination rate); no hyperparameter sensitivity analysis; untested on low-end edge platforms (e.g., Jetson Nano).
4. Unreasonable point cloud similarity metric (prone to error in real scenes); 2D landmark representation loses critical 3D object information; missing key implementation details (e.g., module fusion, coordinate calibration) leading to non-reproducibility; no validation in dynamic/complex scenes (e.g., pedestrians, occlusion, light changes).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors In this paper, the authors propose a low-resource semantic information extraction method based on improved instance segmentation. Authors constructed two specialized datasets for common objects in public indoor areas, designed an instance segmentation mode (MSYOLO) and developed a deep point cloud-based Continuous Frame Dynamic Constraint Method that resolves duplicate object annotations in the map through spatiotemporal consistency verification. The subject of the paper is interesting and up-to-date. In my opinion, authors have to address the following issues: 1) It is not clear to me how the data set collected by the authors described in Section 2 can be used during “geometry mapping” (see Figure 4). 2) Figure 5 - what is CBS? 3) Figure 9 - this is not a diagram but rather a pseudocode. Please rewrite it using indentation in code in order to show where loops, conditions, etc., are. 4) Figure 12, 13 - it is not clear to me how the authors obtained SLAM-like environment maps from data described in Section 2. I believe some description is missing. 5) Tables 2-6 - are teh differences between methods statistically significant? Please discuss it. 6) Without the source codes and dataset, experiments are virtually impossible to reproduce. Please publish source codes in any open repository, for example, GitHub. 7) Because the construction of the dataset is mentioned among the "main contributions" of the paper, it should be made available for readers without making a special request from authors. Please add a link to it in the appropriate section and ensure long-term maintenance.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The authors addressed my remarks. In my opinion, paper can be accepted.