GDCPlace: Geographic Distance Consistent Loss for Visual Place Recognition
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper investigates the geographic distance consistent loss for visual place recognition. After the introduction, the paper presents the related works. Section 3 describes the proposed methods. Section 4 contains the experiments and section 5 contains the discussion. The last section is the conclusion. The paper presents good results, but I have some recommendations:
- Literature review: too many references. Is that much necessary?
- Literature review: please add a table, which summarizes the literature review, for example with the following headers: references, authors, methods, problems.
- Notations and their meanings table is missing
- Test results: Where do the execution results come from? Are they from your own implementation or from others' publications? Execution time? (e.g., Table 1,2)
- Sections 5 and 6 are too brief. Section 5 should include more detailed conclusions. Section 6 should also be more detailed."
- The paper and the appendix contain some test result figures. However, I recommend including more (either in the appendix or on a web platform where the results can be viewed).
- Table A1: Why were these specific parameter values used?
Author Response
Thank you for your thorough review of our manuscript. Please refer to the attached file for details.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper introduces GDCPlace, a new loss function for visual place recognition (VPR), which plays a crucial role in enabling robots and autonomous vehicles to navigate by recognizing locations. The authors mathematically define the Geographic Distance Consistency (GDC) constraint to ensure that similarity scores correspond accurately to geographic distances. They derive an upper bound for the softmax cross-entropy loss under this constraint and propose GDCPlace as the first classification loss function specifically designed for VPR. The study evaluates GDCPlace across 11 various VPR benchmarks, comparing it against different loss functions used in face recognition, image retrieval, and ordinal classification. Please consider the following comments to improve the quality of this paper.
A table summarizing the contributions of related work, including the proposed method, in terms of models, approaches, metrics, simulations, contributions, and limitations would be beneficial.
Additionally, tables and figures should be positioned closer to the relevant text for better readability.
Section 4 would benefit from further analysis of the experimental results.
In Table 6, the preference for the smaller value (51.6) in the Multi-view AM metric should be clarified.
The authors can perform mathematical verification, code verification, cross-validation, generalization testing, convergence analysis, and other validation techniques to ensure the reliability of their results.
Finally, publicly releasing the code would provide a valuable benchmark for future research.
Author Response
Thank you for your thorough review of our manuscript. Please refer to the attached file for details.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsSummary:
A novel loss function designed specifically for Visual Place Recognition (VPR), GDCPlace, is presented in the manuscript. The study makes the case that geographic distance consistency, which is essential for precise location retrieval, is absent from the classification-based VPR techniques now in use. In order to meet this restriction, the authors create the GDCPlace loss function and offer a mathematical description of Geographic Distance Consistency. Extensive experiments across 11 benchmarks demonstrate that GDCPlace outperforms existing classification losses and state-of-the-art (SOTA) methods. In order to improve classification training, the study also investigates Hard Negative Class Mining (HNCM).
General concept comments:
- One important innovation is the adoption of a distance-consistent loss function designed specifically for VPR. The comparison is comprehensive and equitable when compared to various loss functions (e.g., CosFace, ArcFace, LM-Softmax).
- GDCPlace performs better than current losses, however it performs worse in some frontal-view datasets (Table 2). It would be useful to examine the reasons for classification-based losses' difficulties in these situations.
- The application of GDCPlace to ordinal categorisation is covered in the study, however there are no thorough discussions of the practical consequences.
- Further qualitative visualisations of ranking order would support the paper's assertion that GDCPlace achieves geographic distance consistency.
Specific comments:
- Section 3.3 (Line 234): A visualisation of a few chosen hard negatives might be helpful in the description of Hard Negative Class Mining (HNCM).
- Line 471, Table 12: GDCPlace and EigenPlace have similar training efficiency. Could you please clarify if training time is constant for large datasets?
- Visualization of Ranking Errors (Section 4.10, Line 472)
The authors should provide a side-by-side visualisation of ranking errors (i.e., misclassified top-k images) to show where the model fails and how the suggested method could be improved, even though the qualitative results show query image rankings. Misclassified samples may reveal specific failure cases (e.g., viewpoint shifts, extreme lighting conditions) that could guide future improvements.
- Theoretical Justification for Choosing the Sigmoid Function in h(x) (Section 3.4, Line 263)
Although the authors use experimental evidence to support their selection of the sigmoid function for h(x), it would be more persuasive to offer a theoretical justification for why this function is superior to other options (such as piecewise linear functions). Could other functions produce comparable or superior results, and why does the sigmoid function best reflect the geographic distance constraint?
Author Response
Thank you for your thorough review of our manuscript. Please refer to the attached file for details.
Author Response File: Author Response.pdf
