Learning Commonality, Divergence and Variety for Unsupervised Visible-Infrared Person Re-identification
Unsupervised visible-infrared person re-identification (USVI-ReID) aims to match specified people in infrared images to visible images without annotations, and vice versa. USVI-ReID is a challenging yet under-explored task. Most existing methods address the USVI-ReID using cluster-based contrastive learning, which simply employs the cluster center as a representation of a person. However, the cluster center primarily focuses on commonality, overlooking divergence and variety. To address the problem, we propose a Progressive Contrastive Learning with Hard and Dynamic Prototypes method for USVI-ReID. In brief, we generate the hard prototype by selecting the sample with the maximum distance from the cluster center. We theoretically show that the hard prototype is used in the contrastive loss to emphasize divergence. Additionally, instead of rigidly aligning query images to a specific prototype, we generate the dynamic prototype by randomly picking samples within a cluster. The dynamic prototype is used to encourage the variety. Finally, we introduce a progressive learning strategy to gradually shift the model's attention towards divergence and variety, avoiding cluster deterioration. Extensive experiments conducted on the publicly available SYSU-MM01 and RegDB datasets validate the effectiveness of the proposed method.
Paper
Similar papers
Peer review
Summary
This paper proposed a Progressive Contrastive Learning with Multi-Prototype method for USVI-ReID, which learns both shared and diverse information. The Hard Prototype Contrastive Learning method aims to mine distinctive yet significant information, while the Dynamic Prototype Contrastive Learning method is designed to preserve intrinsic diversity features. The experimental results significantly outperform existing methods, validating the effectiveness of the proposed method.
Strengths
1) The paper proposes a straightforward yet effective method, and the experimental results are promising for the USVI-ReID task. 2) The paper is well-written and easy to follow. 3) The proposed method is well-motivated. Its prototype designs enhance the learning of commonality, divergence, and variety. It makes up for the limits of traditional center prototype-based approaches. 4) Figure 2(c) demonstrates that the proposed method generates higher-quality pseudo labels compared to other state-of-the-art methods.
Weaknesses
1) From Table 2, it is evident that the method is based on a relatively strong baseline. Does the paper use different baselines, like ADCA, still achieve good results? 2) The paper introduces a momentum backbone in the PCLHD method but does not fully explain its advantages. 3) Cross-modality associations are especially important for USVI-ReID, what is the association strategy used in this paper? Some details need to be supplemented for clarity.
Questions
What insights does this method offer for other multimodal tasks? How can the approach be adapted to enhance performance in different multimodal scenarios?
Rating
7
Confidence
5
Soundness
4
Presentation
3
Contribution
3
Limitations
Although the method shows significant performance improvement over other unsupervised methods, there is still a considerable gap when compared to fully supervised methods.
Thank you so much for acknowledging the strengths of our method again. We have carefully considered your constructive comments and here are the answers to your concerns. **C1: Do I understand correctly that you do not plan to include a limitation section? If that is the case, I strongly urge you to consider adding such a section.** A1: We greatly value the reviewers' feedback and will be adding a limitation section as shown in the general response to the manuscript. We believe this addition will provide a more complete understanding of our work. **C2: In Eq. 19, Does $v_j$ refer to the query image? Could you please clarify this here? Besides, $v_q$ should be $v_j$ in line 180.** A2: Thank you for your detailed feedback. in Eq. 19, $v_j$ denotes the query image. Besides, We will also correc $y_q$ to$ y_j$ in line 180. **C3: Please note that there is a typo in line 175.** A3: Thank you for pointing out the typo. We will thoroughly proofread the manuscript and correct typos to the best of our efforts.
Thank you for the response, which addressed my concerns.
Summary
This paper proposes a method for unsupervised visible-infrared person re-identification called Progressive Contrastive Learning with Hard and Dynamic Prototype. This method focuses on capturing commonality, divergence, and variety. By introducing hard prototype contrastive learning to emphasize divergence and dynamic prototype contrastive learning to preserve intrinsic feature variety, the approach addresses limitations of existing methods. The proposed progressive learning strategy further enhances model robustness and prevents cluster deterioration. Experimental results on two benchmark datasets, SYSU-MM01 and RegDB, show significant improvements in performance, with notable increases in mAP and Rank-1 accuracy compared to the current state-of-the-art methods.
Strengths
1.The discussion on the limitations of existing methods and how the proposed method addresses these limitations is compelling. The motivation for focusing on commonality, divergence, and variety is well-articulated. 2.The methodology is detailed and well-structured. The explanation of hard prototype contrastive learning, and dynamic prototype contrastive learning is clear. 3.The comparison with state-of-the-art methods is comprehensive. The proposed method’s superior performance. 4.The further analysis, including ARI metric, provides additional insights into the robustness and effectiveness of the method. The analysis of the results is insightful, explaining why the proposed method outperforms others.
Weaknesses
1.While introducing hard prototype contrastive learning can capture divergences, selecting hard samples may lead to model instability during the early training stages. 2.Although a progressive contrastive learning strategy is introduced to mitigate the above problem, balancing the impact of hard samples remains a challenge. 3.The strategy for selecting dynamic prototypes might require more complex implementation and higher computational resources on larger datasets. 4.Although the paper mentions that the code will be released after acceptance, it currently lacks detailed instructions for reproducibility. 5.The accuracy of pseudo-labels directly impacts the model training. Some parameters in DBSCAN are very important and should be analyzed further.
Questions
1.Is L_CPCL involved in the optimization process during the second stage? 2.Is the comparison fair with GUR without camera information?
Rating
7
Confidence
4
Soundness
3
Presentation
4
Contribution
3
Limitations
The method relies on the DBSCAN clustering algorithm to generate pseudo-labels. For large-scale datasets, DBSCAN might not perform well, thereby affecting the overall method's performance.
Dear Reviewers,\ Thanks very much for your time and valuable comments. We understand you're busy. But as the window for responding and paper revision is closing, would you mind checking our response and confirm whether you have any further questions? We are very glad to provide answers and revisions to your further questions.\ Best regards and thanks,\ Authors of # 1045
Summary
The paper is presented in the field of unsupervised visible-infrared person re-identification (USVI-ReID) which aims to match individuals in infrared images to those in visible images without annotations. Existing methods often use cluster-based contrastive learning, which fails to account for divergencies by focusing solely on commonalities. To overcome this, the authors propose a Progressive Contrastive Learning method with Hard and Dynamic Prototypes, which emphasizes divergencies and feature variety through the use of hard prototypes and a progressive learning strategy.
Strengths
The motivation behind the proposed methodology is exceptionally clear, as is the well-structured solution scheme. The logic of the presentation is sound, and the progression of the proposed solution is easy to follow, making the entire approach both coherent and engaging. The innovative idea of adopting contrastive learning to enhance discrimination by incorporating diversity information is particularly noteworthy. This approach stands out as it not only addresses existing challenges but also brings a fresh perspective to the field. The results, which have been demonstrated on two benchmark datasets, are truly impressive. They outperform recent state-of-the-art solutions that have been presented in top-tier venues in the computer vision (CV) field over the past few years. This achievement highlights the effectiveness and robustness of the proposed methodology.
Weaknesses
The concept of farthest equal diversity in the HPCL (equations 6 and 7) is based on a strong assumption. Distance is a weak measure and a poor indicator of how the samples within a cluster are actually spread out. While variance is a simple metric, it captures the statistical distribution of the samples within the cluster more effectively. The DPCL section is not clearly articulated. Specifically, the concept of the query and its application in equation 19 are obscure and require further clarification. The loss function L_PCLHD relies on the constant E_CPCL , which is arbitrarily set to 50 without any justification or ablation study to support this choice. From the results presented, it appears that all the ablation studies commence with 50 epochs of CPCL. Additionally, although it is common practice to integrate the proposed methodology into unsupervised solutions, I believe that knowing the number of classes in advance and setting the number of clusters accordingly constitutes a form of supervision. While it is less intensive than labeling every sample with its respective class, it is still a form of supervision. In my opinion, these methods should be categorized under "unlabeled samples" rather than "unsupervised" learning.
Questions
Did the authors explore alternative methods for representing diversity aside from using the farthest sample? If such exploration was undertaken, it would be beneficial to discuss the findings. Conversely, if no alternative methods were considered, does this represent a limitation of the proposed solution? Understanding the rationale behind the chosen approach or the constraints of the methodology would add valuable context to the study. How does the performance of the proposed solution vary with different values of the E_CPCL constant? Was any ablation study conducted to investigate this aspect? An examination of the impact of varying E_CPCL on the solution's effectiveness would provide deeper insights into the robustness and adaptability of the proposed method. Such analysis would be crucial for validating the consistency and reliability of the results.
Rating
4
Confidence
4
Soundness
3
Presentation
2
Contribution
3
Limitations
The limitations of the proposed approach are not adequately addressed. The weaknesses related to assumptions (e.g. diversity) and constant values definitions, should have brought the authors to a deeper ablation study and thus comprehension of the limitations of the proposed solution.
Thank you once again for providing feedback on our paper and for your willingness to consider improving the score. **C1: I still have doubts about using distance in HPCL since in many Re-ID solution there is significant effort also in metric learning showing that distance is not capturing all the characteristics of the feature space.**\ A1: We agree with your observation that distance is not a perfect method for capturing all the characteristics of the feature space. **However, the key idea of HPCL is to focus the network on divergence using distance-based hard prototype learning, rather than on measurement methods.** We have used the simple distance measure to verify that focusing on divergence is important. We also acknowledge that there are more complex ways to measure divergence, and we believe that using more advanced metrics could improve our approach. **C2: Concerning the $E_{CPCL}$ the authors confirm the necessity of ablation on such a parameters since 40 is a better value. What is missing is a comment on why such a parameters has such an impact and with such characteristics.**\ A2: The parameter $E_{CPCL}$ is crucial in determining the timing for focusing on divergence and variety. When $E_{CPCL}$ is set to 20, the hard prototypes are primarily composed of noisy samples. However, when $E_{CPCL}$ is set to 40, these prototypes are more likely to consist of clear hard samples. For instance, at the 20th epoch, the ARI score for IR image pseudo-label accuracy is only 0.68, but it improves to 0.89 at the 40th epoch, which demonstrates how $E_{CPCL}$ affects the quality and composition of prototypes. When $E_{CPCL}$ is set beyond 40, the network has already somewhat converged, which reduces the impact of proposed method.
Dear Reviewer mgXA, Thank you for your previous feedback and for mentioning your intention to raise the score. We noticed that the score has not been updated. As the discussion phase comes to an end, we want to ensure that we have thoroughly addressed all of your concerns. If there are any additional points of concern that you believe would contribute to the improvement of our manuscript, please let us know. We genuinely appreciate your expertise and the dedicated time you have devoted to reviewing our work. Warmest regards, Authors
Summary
The paper proposes a method for unsupervised visible-infrared person ReID. Towards this goal, it extends the PGM method with a progressive contrastive learning with hard and dynamic prototypes. Initially a typical centroid prototype contrastive learning approach is used. A mixed batch of images is encoded through an online encoder and contrastive learning is used to update this encoder using contrastive learning with the centroid prototypes computed by DBSCAN clustering the training data and updating it with a momentum-updating strategy throughout the training. The paper hypothesizes that additionally hard samples should be used as prototypes to find divergent samples and dynamic prototypes should be used to preserve variability during training. Hard prototypes are selected as those samples within a cluster that are farthest away from the centroid prototypes and dynamic prototypes are random samples from a cluster and encodes them with an encoder that keeps a momentum version of the online weights. Since the hard and dynamic prototypes might not be well-defined at the start of training, the proposed method initially uses the standard centroid prototype contrastive learning and after a certain period of training, changes to a loss function based on hard and dynamic prototype contrastive learning. This setup shows significant improvements of the PGM baseline, but also over the current state of the art.
Strengths
- The paper is fairly easy to follow. - The results look strong. - Code is promised to be released upon acceptance. - Overall the idea makes sense and the ablations do show that the different parts of the method are relevant.
Weaknesses
- L215 states that each batch contains 16 person IDs and 16 samples per person. Is this a confusion and it should be based on the assignments from the clustering? Or are these actually person IDs from the annotations? This is kind of a big point, so please make sure to answer this in the rebuttal. I checked some other papers and there similar things are mentioned, but if this is truly based on annotations, the approach CANNOT be called unsupervised, unless you show you can do without annotations here! - The novelty isn't that great, the base method PGM covers most of the approach and the real novelty lies in a slight modification to the contrastive learning setup for a part of the training. The dynamic prototype learning, as the paper states, is inspired by MoCo and when searching a bit for hard examples in contrastive learning, there is also quite some literature available. Nevertheless, in this specific setting, there is a contribution of showing the improvements are possible. - I find the related work very weak! The paper is completely focused on the contrastive learning aspect of the PGM method, however, the related work section is purely focused on Visible-Infrared and unsupervised person ReID. Why is there no discussion about unsupervised learning in general, specifically with a focus on contrastive learning and its modifications?! - While the performance boost in general is quite good, figure 2a does raise the question how stable it is w.r.t to some of the hyperparameters. Then again a fixed value of 0.5 isn't too unexpected and I guess it might have even been an initial default value. Nevertheless an actual validation set for these kinds of ablations would have been valuable, but I guess also uncommon in the community. - Overall the discussion of the results is often a bit meaningless and just describes the values in the tables, partially even in a bit of a weird fashion. For example L238-240 points out that the proposed method "remarkably" is able to outperform even some semi-supervised methods, but quite some other unsupervised methods in the table also do so. Then in L241-246 it's pointed out that even some supervised methods from 2021 are outperformed, but almost every other supervised method outperforms every single unsupervised method in the table.
Questions
- Why are the "Baseline" results in table 2 so different from the PGM results? I get the fluctuation of ~1% for All Search, but why is Indoor Search so much better in your training setup? - Why did you duplicate almost every single equation just for visible and infrared? This is a major waste of space and could have been used for actually relevant discussions. - In figure 2c, it very much confused me to see a line between the data points since the different methods are in no way connected. - Is figure 3 generated based on training of test images?
Rating
5
Confidence
4
Soundness
3
Presentation
2
Contribution
3
Limitations
The authors claim "We have not found the limitations of the work." but then again the paper is far from solving the problem. Given that this indicates there are obviously limitations, I find it somewhat lazy to not put some effort into coming up with a meaningful discussion of what still needs to be improved. And then this is actually outrageous: "There is no societal impact of the work performed." The complete paper is about person re-identification, now even beyond the visible spectrum. Performing this kind of research can have major effects on how effective surveillance is and it's ridiculous to state that there is no impact! I will not judge someone for doing this research out of scientific curiosity, but thinking it has no impact and propagating that belief is not acceptable.
We sincerely thank you once again for your valuable comments. We have carefully considered your concerns and have provided our responses below. **C1: Could you actually elaborate on W3? You plan to add additional related work, but maybe give a list of the papers you plan to discuss here? As is, I can only hope you will come up with a meaningful discussion and that doesn't feel very convincing.** A1: Thank you for your follow-up question. We plan to additional related work about unsupervised learning in general, specifically with a focus on contrastive learning and its modifications. Specifically, we will discuss the following related works:\ Contrastive learning has achieved remarkable advancements in visual representation learning by pulling positive samples closer together and pushing negative samples further apart in feature space[1,2,6]. In contrastive learning, positive and negative samples play a crucial role and have a direct impact on the model's performance. Therefore, some data augmentation methods[2,3] have been proposed to augment positive pairs to learn discriminative features. For negative pairs, which have received less attention in contrastive learning, MoCo[1,4,5] series construct a memory bank to maintain numerous negative samples for increasing model’s performance. However, these methods define samples from different instances as negative, which is not feasible for ReID. In ReID datasets, many samples have the same identity, causing many positive samples to be wrongly labeled as negative.\ [1]Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. CVPR, 2020.\ [2]Xuyang Zhao, Tianqi Du, Yisen Wang, Jun Yao, and Weiran Huang. ArCL: enhancing contrastive learning with augmentation-robust representations. ICLR, 2023.\ [3]Demirel B U, Holz C. Finding order in chaos: A novel data augmentation method for time series in contrastive learning. NeurIPS, 2023.\ [4]Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint, 2020.\ [5]Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. ICCV, 2021.\ [6]Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liqun Chen, Belinda Zeng, Trishul Chilimbi. Vision-language pre-training with triple contrastive learning. CVPR, 2022. \ **If you have any more relevant works in mind, we would greatly appreciate your suggestions.** **C2: Also, as another reviewer highlighted, will you be adding a limitations section to the paper?**\ A2: We greatly value the reviewers' feedback and will be adding a limitation section as shown in the general response to the manuscript. We believe this addition will provide a more complete understanding of our work. **C3: You have promised to add quite some things to the final paper and I'm wondering how all of this will fit and which parts will be sacrificed and moved to the supplementary to make all of this fit.**\ A3:To accommodate the additional content, we plan to move the discussion of Unsupervised Single-Modality Person ReID in the related work section, as well as Figure 2a and its related discussion, to the supplementary material. This will allow us to maintain the core content of the paper while ensuring that all important details are still accessible. **C4: In another comment you wrote "For instance, in image-text matching tasks, hard and dynamic prototypes help capture divergence and variety in visual and textual data, leading to better alignment between the two modalities." Is this some idea that you came up with, or is there an actual paper that does this?**\ A4: This idea, to the best of our knowledge, has not appeared in any existing papers. It is an original concept that we plan to explore further as we expand this work.
Dear Reviewer edeS, We sincerely appreciate your insightful review, as it plays a pivotal role in enhancing the quality of our manuscript. Furthermore, we have carefully considered your feedback to refine our paper further. As the discussion phase comes to an end, we want to ensure that we have thoroughly addressed all of your concerns. If there are any additional points of concern that you believe would contribute to the improvement of our manuscript, please let us know. We genuinely appreciate your expertise and the dedicated time you have devoted to reviewing our work. Warmest regards, Authors
The authors have successfully replied to some of my concerns. I still have doubts about using distance in HPCL since in many Re-ID solution there is significant effort also in metric learning showing that distance is not capturing all the characteristics of the feature space. Anyhow the authors show that simple distance is better than variance. I appreciate the effort. Concerning the E_{CPCL} the authors confirm the necessity of ablation on such a parameters since 40 is a better value. What is missing is a comment on why such a parameters has such an impact and with such characteristics. With the rebuttal I would like to raise my score to borderline accept.
Thank you very much for addressing most of my concerns. Considering the other reviews, I believe the revised manuscript will be nice, and I am willing to raise my score. However, after re-reading the manuscript, I have few new questions and suggestions: (1) Do I understand correctly that you do not plan to include a limitations section? If that is the case, I strongly urge you to consider adding such a section. (2) In Eq. 19, Does $v_j$ refer to the query image? Could you please clarify this here? Besides, $y_q$ should be $y_j$ in line 180. (3) Please note that there is a typo in line 175.
Could you actually elaborate on W3? You plan to add additional related work, but maybe give a list of the papers you plan to discuss here? As is, I can only hope you will come up with a meaningful discussion and that doesn't feel very convincing. Also, as another reviewer highlighted, will you be adding a limitations section to the paper? You have promised to add quite some things to the final paper and I'm wondering how all of this will fit and which parts will be sacrificed and moved to the supplementary to make all of this fit. Finally, in another comment you wrote "For instance, in image-text matching tasks, hard and dynamic prototypes help capture divergence and variety in visual and textual data, leading to better alignment between the two modalities." Is this some idea that you came up with, or is there an actual paper that does this?
Thank you for your feedback. We understand your concerns, which reflect broader issues that exist across all AI technologies. From a dialectical perspective, AI can be used for both good and ill. However, as the renowned scholar Fei-Fei Li has also stated, “I believe that humanity can effectively address the challenges posed by AI, as long as we adhere to human-centered artificial intelligence.” Besides, the question of whether person re-identification (ReID) is inherently guilty is beyond the scope of this paper, which focuses on the technology itself rather than policy and regulation.\ Here are the answers to your concerns.\ **C1: The work in this paper is akin to performing research on improving facial recognition, as such, my position is that it should not be published.** \ A1: ReID offers greater privacy compared to facial recognition technology. ReID uses structured data instead of personal photos. Specifically, it uses feature vectors from images at a given time, represented as numerical values. Besides, this technology only matches the same individual and cannot identify specific identities.\ **C2: Now considering where infrared technology is most often used (surveillance and military purposes) it seems questionable, at best whether such work should be published at scientific and academic venues. In my opinion such work has no home in either scientific or academic venues.**\ A2: ReID has received significant attention in computer vision for a long time. Many significant ReID studies [1-7] have been published in NeurIPS. ReID cannot be published, many published papers in this field would also need to face similar scrutiny. \ [1] HAP: Structure-Aware Masked Image Modeling for Human-Centric Perception. Junkun Yuan et al. NeruIPS 2023.\ [2] TransMatcher: Deep Image Matching Through Transformers for Generalizable Person Re-identification. Shengcai Liao et al. NeruIPS 2021.\ [3] Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID. Yixiao Ge et al. NeruIPS 2020.\ [4] The Dilemma of TriHard Loss and an Element-Weighted TriHard Loss for Person Re-Identification. Yihao Lv et al. NeruIPS 2020.\ [5] Learning Disentangled Representation for Robust Person Re-identification. Chanho Eom et al. NeruIPS 2019.\ [6] FD-GAN: Pose-guided Feature Distilling GAN for Robust Person Re-identification. Yixiao Ge et al. NeruIPS 2018.\ [7] Deep Neural Networks with Inexact Matching for Person Re-Identification. Arulkumar Subramaniam et al. NeruIPS 2016.\ **C3: The purpose largely should be to the benefit of society, and (again in my opinion) it is hard to justify how surveillance infrastructures or military applications have improvements for society, narrowly or broadly.**\ A3: It’s important to clarify that our research has always been aimed at using this technology for benefit society, not for military or harmful applications. For example: 1) Helping locate missing persons, such as lost children in public spaces. 2) Enhancing customer behavior analysis, optimizing store layouts, and improving personalized services. 3) Advancing AR experiences through precise tracking of individuals and interactive content integration.\ **C4: The authors do not discuss the limitations of their method or the social implications.**\ A4: We greatly value all reviewers' feedback and will be adding a limitation section as shown in the general response to the manuscript. We believe this addition will provide a more complete understanding of our work.\ **C5: The authors do not discuss the social implications.**\ A5:We acknowledge that any technology can be misused, but we believe risks of ReID can be managed. Specifically, tracking individuals in public spaces with ReID technology could raise privacy concerns. However, public surveillance systems using ReID are typically controlled by authorized institutions, which reduces the risk of misuse.\ **C6: The technology itself lends directly to surveillance, weaponization, and breaches of human rights.**\ A6: As ethics reviewer fgoD noted “This method is largely technical, and is developed through experiments on publicly available datasets. As such, the work developed in this paper does not raise direct issues of ethical, safety, or environmental concerns. No creatures were harmed in the making of this research.” In fact, Our research does not raise direct issues of ethical, safety, or environmental concerns.
Thank you for your feedback. We appreciate your acknowledgment that our method is developed through experiments on publicly available datasets and does not raise direct issues of ethical, safety, or environmental concerns.\ We acknowledge that any technology can be misused, but we believe risks of ReID can be managed. For example, tracking individuals in public spaces with ReID technology could raise privacy concerns. However, public surveillance systems using ReID are typically controlled by authorized institutions, which reduces the risk of misuse. We will add a discussion on the ethical and social implications in future revisions of our paper to better align with NeurIPS guidelines.
After reviewing the rebuttal carefully, the authors have addressed my previous concerns in detail. I have also considered the opinions of other reviewers:\ (1)I agree with the other reviewers that the authors should add the limitations discussed in the general rebuttal to the original manuscript.\ (2)I agree more with Ethics Reviewer fgoD than with Ethics Reviewer zWHs. I argue this work is largely technical and does not present direct ethical concerns.\ Considering the opinions of other reviewers, I have decided to increase the score from 7 to 8.
We sincerely appreciate your response and for your willingness to improve the score.
Response to Ethics Review Concerns
Dear Area Chairs and Reviewers The recommendation from Ethics Reviewer #ZWH demonstrates a serious bias against surveillance technologies. We contend that technology itself is neutral, and its impact is determined by its specific use case and management policies. It is widely accepted that the primary purpose of technologies is to benefit society and promote progress. Furthermore, many artificial intelligence technologies, particularly in computer vision, are ultimately applied in surveillance systems. This should not negate the valuable contributions made by researchers in these fields. As the Ethics Review Chairs stated, “ethics reviews are meant to be educational, not prohibitive. Our goal is not to police submissions, but instead to prompt reflection.” Therefore, we sincerely hope that the Area Chairs and Reviewers will not consider this biased opinion, to ensure a fair and objective evaluation of the paper’s technical and scientific contributions. For a detailed discussion of Ethics Reviewer #ZWH’s other comments, please refer to the reply for Ethics Reviewer #ZWH. Warmest regards, Authors
Decision
Accept (poster)