Skip to Content
MDPIMDPI

Advanced

by
  • Xinxing Tian,
  • Zhilin He and
  • Tianhao Gu *
  • et al.

Reviewer 1: Po-Chyi Su Reviewer 2: Anonymous Reviewer 3: Anonymous

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors
  1. The following sentence from the abstract is misleading and requires revision:
    “This paper proposes a lightweight self-supervised framework from single RGB images for multi-scale feature extraction and artifact elimination in monocular depth h estimaetion (MsGf).” First, the phrase “self-supervised framework from single RGB images” is unclear and potentially incorrect. Second, artifact elimination is not typically a core component of monocular depth estimation and may be misleading in this context. Lastly, “estimaetion” is a typographical error and should be corrected.
  2. The proposed components, CDMs and SEGF, appear to be heuristic and lack strong novelty. While it is understood that deep learning approaches often involve heuristic designs and require ablation studies to validate their effectiveness, it is unclear whether the use of multiscale features and edge-guided filtering can still be considered significant contributions in the current state of the field.
  3. If the goal is depth estimation, deep learning methods typically achieve accurate results by learning directly from noisy data. Artifact suppression is usually intended to enhance visual quality, not necessarily to improve estimation accuracy. Altering the input data prior to estimation may risk compromising its fidelity.
  4. In the ablation study, the "Model" variant uses only the basic network structure without any additional modules or mechanisms. Why, then, does it have the largest number of parameters? Besides, “BA loss” is not defined. The “loss” in Table 4 should be explained.
  5. It is not easy to tell the difference in Figure 5.
  6. Further proofreading is required.

Author Response

Please see the atachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Please check the document

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper proposed a lightweight self-supervised monocular depth estimation framework. Cross-dimensional multi-scale feature extraction modules were designed to reduce parameters. Additionally, a Sobel edge perception guided filtering module was developed to capture structural and edge features. Experimental results showed that the proposed approach achieved higher accuracy and smaller model size compared to several existing approaches. Although this paper demonstrates innovations in model design, there are still several issues that need to be addressed before the paper can be considered for publication:

  1. Related work section: Please include several recent studies that were published in 2025.
  2. Section 3.1: The introduction discusses the limitations of CNN- and transformer-based methods. CNN-based methods, in particular, struggle with weak-texture regions such as smooth walls and sky areas. Since the proposed framework employs depth-wise separable convolutions for feature extraction, please clarify whether and how the method addresses these challenges.
  3. Figure 2: The architecture requires more detailed explanation. For example, how were the three feature scales concatenated despite differing sizes? What are Kernel Generator and GF blocks? Additionally, the inputs to the three upsampling blocks appear incorrect. Does the subfigure labeled “UpConv Block” represent the “Upsampling Block”?
  4. Equation 2: According to Figure 3a and Lines 174-176, both GELU and an additional 1×1 Conv were used. Please revise accordingly.
  5. Equation 5 and Figure 3b are inconsistent. The 1×1 Conv shown in the equation is not reflected in Figure 3b. Moreover, according to Equation 5 and Line 192, the outputs of the CDMs should not serve as inputs to the filtering module. Please check them carefully.
  6. Table 1: The citation for the Lite-Mono series is incorrect.
  7. Figure 4: It is recommended to add an extra row with ground truth depth maps for clearer comparison.
  8. Table 3 compares more models (e.g., MViT-Depth series and LiteGfm) than Table 1. For consistency, please also include the results of these models in Table 1.
  9. Section 4.4: Please specify the architecture of Model 1. Which modules were used for feature extraction?
  10. Figure 6: Why do the feature maps of LiteGfm appear smoother compared to the proposed model?
  11. Discussion section: The limitations of the proposed framework should be analyzed in more depth, ideally with examples of failure cases.
  12. It is strongly recommended that the authors make the source code publicly available to ensure reproducibility.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed the issued listed in our previous review.

The authors may add the ground truth of depth information in Figs. 1 and 6 so that we can know which result is closest to the ground truth.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have adequately addressed most of my comments in the revised manuscript. The paper can be accepted after the following revisions:

  1. Figure 2: Please double-check the inputs of the three upsampling blocks. In the current figure, features from stage 3 and stage 2 are shown as inputs to the right block, while features from stage 1 are connected to the middle block. I believe that features from stage 2 should instead be connected to the middle block, and features from stage 1 should be connected to the left block.
  2. If dense depth maps are not available in the KITTI dataset, please clarify how the authors handled pixels in the RGB images without corresponding depth values when training the model and calculating evaluation metrics. Were these pixels simply ignored? Additionally, please clarify whether the sparse ground truth is sufficiently reliable for evaluation.
  3. Table 1: After including the results of the MViT-Depth series and LiteGfm, the performance improvement appears to be minor, particularly compared to LiteGfm. Please revise Section 4.2.1 (Quantitative Comparison) accordingly and explicitly include the comparison with both the MViT-Depth series and LiteGfm.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf