For academic editor:

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming.

Response: We thank the editor for the reminder. We have carefully checked and revised our manuscript to ensure full compliance with PLOS ONE’s style and formatting requirements, including the file naming conventions and all other journal specifications. All necessary adjustments have been made accordingly.

2. Please note that PLOS One has specific guidelines on code sharing for submissions in which author-generated code underpins the findings in the manuscript.

Response: We agree with the suggestion on making the code available. At present, we provided an initial version of the code for review and verification purpose, with the url: https://github.com/Feynluo/uamp-sam2 . We will further organize, refine, and fully release the complete code on a public repository upon formal acceptance of the manuscript.

3. Thank you for stating the following financial disclosure: This research was funded by Shaanxi University of Technology, grant number SLGRC202416. Please state what role the funders took in the study.

Response: The funder (Shaanxi University of Technology, grant number SLGRC202416) provided financial support only for this study. The funder had no role in the design of the study, data collection, data analysis, interpretation of results, writing of the manuscript, or decision to publish. And the role of funder statement is amended in the cover letter.

4. Thank you for stating the following in the Acknowledgments Section of your manuscript: This research was funded by Shaanxi University of Technology, grant number SLGRC202416. Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Please include your amended statements within your cover letter; we will change the online submission form on your behalf.

Response: We have removed the funding-related text from the Acknowledgments section, and the funding statement: This research was funded by Shaanxi University of Technology, grant number SLGRC202416, is included in the cover letter.

5. We note that Figure(s) 1, 2, 3, 5, 6, 7, 8, 9, 10, 11 in your submission contain copyrighted images. All PLOS content is published under the Creative Commons Attribution License (CC BY 4.0), which means that the manuscript, images, and Supporting Information files will be freely available online, and any third party is permitted to access, download, copy, distribute, and use these materials in any way, even commercially, with proper attribution. For more information, see our copyright guidelines: http://journals.plos.org/plosone/s/licenses-and-copyright.

Response: We sincerely appreciate the reminder concerning the copyright and licensing issues of the figures in our manuscript.

We hereby confirm that all figures included in this manuscript are original works created by the authors and do not involve any third-party materials, copyrighted content, or previously published images, and none of the figures pose conflicts with the CC BY 4.0 licensing terms.

And we provide the original figures created with draw.io tool in the supplement materials.

6. If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise.

Response: We thank the editor for the reminder concerning the citations suggested by the reviewers. We have carefully reviewed and evaluated all the recommended publications for their relevance to the present study. Those that are relevant and provide valuable support to our work have been appropriately cited in the revised manuscript. In accordance with the editorial guidance, we have not cited publications that are not directly relevant to the content of this paper.

Reviewer #1:

1. The abstract and conclusion need to be improved. The abstract must be a concise yet comprehensive reflection of what is in your paper. Please modify the abstract according to “motivation, description, results and conclusion” parts. I suggest extending the conclusions section to focus on the results you get, the method you propose, and their significance.

Response: We sincerely appreciate the valuable comments.

The abstract has been thoroughly revised to concisely and comprehensively reflect the manuscript, following the "motivation, description, results and conclusion" structure. The motivation clarifies SAM 2’s limitations in long-term VOS (error accumulation, poor occlusion robustness, inadequate adaptation to dynamic objects). The description details our proposed UAMP (a SAM 2 variant) and its core uncertainty-aware memory propagation framework with three key components. The results present key experimental findings, including UAMP’s superior performance across five VOS benchmarks, a maximum 5.6-point J&F improvement over SOTA methods, and stable performance in complex scenarios. The abstract’s conclusion summarizes UAMP’s effectiveness in enhancing SAM 2 for practical VOS.

The conclusion has been extended and optimized to focus on the proposed method, results, and their significance. We clarified UAMP’s core innovation and its targeted design to address SAM 2’s limitations. Key experimental results are elaborated, confirming UAMP’s effectiveness in mitigating error accumulation and adapting to dynamic objects, which provides a reusable module for VOS.

All revisions have been made in the revision.

2. The Section 1 and Section 2 is too short. I suggest the authors can merge section 1 and section 2.

Response: Thanks for your valuable suggestions. Section 1 (Introduction) focuses on the research background, SAM 2’s limitations in long-term VOS, the motivation of our UAMP method, and a brief overview of its core ideas, aiming to clarify the research gap and the necessity of our study. In contrast, Section 2 (Related Works) provides the related research, summarizing the development of VOS, and different paradigms, including the Memory-based VOS and Consistency Modeling, and highlighting the differences between these methods and our UAMP, helping readers grasp current research progress and our study’s innovation.

And further, according the following advices of 3 and 5, we further improve the introduction and related works section in the revision.

3. What is the motivation of the proposed method? The details of motivation and innovations are important for potential readers and journals. Please add this detailed description in the last paragraph in section I. Please modify the paragraph according to "For this paper, the main contributions are as follows: (1) ......" to Section I. Please give the details of motivations. In Section 1, I suggest the authors can amend your contributions of manuscript in the last of Section 1.

Response: Thanks for the suggestions. First, detailed descriptions of the motivation for the proposed UAMP method have been added to the last paragraph of Section I. The expanded motivation clarifies the research background (the wide application of SAM 2 as a foundational VOS model) and elaborates on the critical limitations of SAM 2 in long-term VOS tasks—specifically, error accumulation caused by its greedy-selection memory architecture, insufficient robustness to object occlusions and reappearances, and inadequate adaptation to dynamic appearance and motion variations of objects in complex scenarios.

Second, the contributions of the manuscript have been revised and placed at the end of Section I, following the format suggested by the reviewer: "For this paper, the main contributions are as follows: In summary, our main contributions are as follows: 1) We propose UAMP, a consistent video object segmentation model, which is specifically designed to address the inherent lim-itations of SAM 2 in long-term VOS tasks, with a core uncertainty-aware memory propagation framework, providing a reusable module for optimizing memory-based video segmentation models. 2) We introduce an uncertainty-aware memory propagation mechanism via re-gion-adaptive memory fusion by appearance and motion uncertainty estimation to refine the feature representation. At the meantime, a long-term memory updating mechanism and score-based short-term memory selection is designed to reduce the disturbance in crowded scenes by a mixture of motion and affinity scores. 3) Our UAMP achieves state-of-the-art performance on SA-V, MOSE, LVOSv1 and LVOSv2, and other VOT benchmarks, demonstrating effectiveness across diverse datasets".

4. The description of manuscript is very important for potential reader and other researchers. I encourage the authors to have their manuscript proof-edited by a native English speaker to enhance the level of paper presentation. There are some occasional grammatical problems within the text. It may need the attention of someone fluent in English language to enhance the readability.

Response: Thanks for the valuable suggestion regarding the manuscript’s presentation and English language readability. Specifically, we have invited a native English speaker with expertise in academic writing (focusing on computer vision and machine learning fields) to conduct a comprehensive proof-editing of the entire manuscript. And the proof-editing work has been completed in the revised manuscript.

5. The introduction section of the paper needs to revise according to the timeline of technology development. Please update references with recent paper in CVPR, ICCV, ECCV et al and Elsevier, Springer. In your section 1 and section 2, I suggest the authors amend several related literatures and corresponding references in recent years. For example: Dual Degradation Image Inpainting Method via Adaptive Feature Fusion and U-Net Network (Applied Soft Computing); CAAT: Image Super-resolution Algorithm via Channel Attention and Transformer (Array); MGNet: RGBT tracking via cross-modality cross-region mutual guidance (Neural Networks); Crack segmentation network via difference convolution-based encoder and hybrid CNN-Mamba multi-scale attention (Pattern Recognition)

Response: Thanks for the valuable advice. And the introduction section of the paper has been revised according to the timeline of technology development with the updated recent papers. Further, in section 1 and section 2, new related literatures have been amended in the revision.

Specifically, “SAM2Long [4] aimed to address the failure of SAM 2’s FIFO fixed-window memory mechanism in handling target reappearance by introducing a hierarchical memory tree structure for flexible memory retrieval and management, though it suffered from increased memory overhead, potential retrieval latency, and failed to resolve the confusion between simi-lar-looking instances during memory retrieval. SAMURAI [3] integrated Kalman filtering into SAM 2’s memory update process to model target motion trajectories and suppress error ac-cumulation through a prediction-update mechanism, yet its performance was heavily de-pendent on motion prediction accuracy in fast-changing scenarios, and it could not effectively handle severe target deformation that led to large prediction errors. DAM4SAM [5], proposes a novel distractor-aware memory (DAM) module for the SAM 2 model, which is composed of recent appearance memory (RAM) and designed to address the interference of distractor objects (irrelevant objects with similar appearance to the target) in VOS tasks. However, DAM4SAM over-relies on the recent appearance features stored in RAM, making it less adaptable to scenarios where the target undergoes significant appearance changes over time. DC-SAM [45] leverages multi-source features and generates positive and negative prompts by ensuring prompt consistency, integrating with SAM/SAM2 to achieve in-context segmentation for both images and videos. and improves the model's adaptability to complex scenes. But it focuses on prompt-tuning adaption, and lacks effective long-term temporal correlation modeling capabilities, making it difficult to maintain consistent segmentation performance when the target undergoes continuous appearance changes (such as scale variation, deformation) in long video sequences”.

6. Please give the details of proposed method for proposed model. I suggest the authors amend the calculation of your size of proposed method and the details is important for proposed method.

Response: We sincerely appreciate your valuable comment. Following your suggestion, we have supplemented the detailed implementation and dimensional calculation of each core module of the proposed model, and revised the relevant content in the section 3 and section 4.12 of model details.

Specifically, “For image encoder, the input image is first scaled to the resolution of 1024, and the image embeddings are obtained by fusing the stride 16 and 32 features from Stages 3 and 4 of the Hiera [25] image encoder with the dimension 448×64×64 and 896×32×32 respectively. These multi-scale features are then fused via a pyramid network following [1], and model dimension of Hiera is 256. The positional embeddings are calculated with the windowed absolute positional embeddings, and the positional information spanning across windows is obtained by interpolating the positional embeddings. For memory attention, the input feature map is the 256×64×64 embeddings generated by the memory encoder, the 2d spatial Rotary Positional Embedding (RoPE) [30] with the dimension of 256 is utilized in both self-attention and cross-attention layers, and this module composes 4 layers. For mask decoder, the stride 4 and 8 features from Stages 1 and 2 are added to the up-sampling layers to get high-resolution features. Following [6], the appearance uncertainty prediction module comprises one 1×1 convolution and two 3×3 convolutions. It takes the input of a concatenation of current frame feature, last frame feature, and last mask prediction, and outputs a one-channel change probability mask of the query token. For memory bank, the memory features 256×64×64 are projected to the dimension of 2048 for 8 memory banks via a linear layer, and the short-term memory selection is designed follows DAM [5] as discussed in Section 3.3. The long-term memory updating composes 4 layers of the transformer-like structure, in which the self-attention layer is replaced with the linear-attention layer. And the matrix state is projected to the same dimension as short-term memory with a linear projection layer.”

7. The content of experiments needs to amend related experiments to compare related SOTA in recent three years. I recommend the authors amend related experimental results of proposed method of SOTA according to the published paper in IEEE, Springer and Elsevier.

Response: We sincerely appreciate your comment. We have amended related experimental results with the state-of-the-art (SOTA) methods in the field published in the past three years in the Table 1 and Table 2 of section 4.2.

Specifically, we compare with the Cutie-bas [20] method presented at CVPR 2024, LiVOS [46] at CVPR 2025, SAM2.1[1] at ICLR 2025, and SAM2.1Long [4] at ICCV 2025. All the selected approaches provide complete open-source implementations or detailed experimental protocols in their original papers, which ensures the fairness of our comparative evaluation.

8. However, the manuscript, in its present form, contains several weaknesses. Adequate revisions to the following points should be undertaken in order to justify recommendation for publication.

Response: We sincerely appreciate the reviewer for the constructive comments and valuable suggestions. We fully agree that the manuscript needs further improvement to enhance its quality. We have carefully revised the manuscript according to the comments, and the detailed revisions are addressed point by point in the revision.

9. In the conclusion section, the limitations of this study and suggested improvements of this work should be highlighted.

Response: We sincerely appreciate the reviewer’s constructive suggestion. Following this comment, we have supplemente

Attachments

Attachment

Submitted filename: Response to Reviewers.docx

For Editor:

Required revisions:

1) Clarification and formal definition of uncertainty modules

The manuscript should clearly define the appearance and motion uncertainty components and describe how these quantities are supervised or learned during training. The current description remains partly conceptual and does not fully specify the mathematical formulation.

Response: Thank you for your careful and constructive comment. We have further strengthened the mathematical definition and learning mechanism of appearance uncertainty and motion uncertainty to avoid conceptual ambiguity.

Specifically, we have clearly defined the formulation of appearance uncertainty in Eq (5) and its corresponding supervised loss in Eq (6), which explicitly guides the learning of appearance reliability during training. For motion uncertainty, we have provided a precise mathematical formulation in Eq (8-9). In the revised manuscript, we have also added detailed explanations for each component, clarifying how these uncertainty values are modeled, optimized, and integrated into the overall training pipeline. All components are now fully specified with explicit mathematical expressions, ensuring the description is concrete and reproducible.

2) Reproducibility of the proposed framework

Several key elements of the method require clearer specification to ensure reproducibility, including:

a) the gating mechanism used for balancing long-term and short-term memory

b) the exact formulation of the uncertainty estimation modules

c) the loss functions and weighting schemes used during training

d) the full specification of the Kalman filter used for motion scoring

Response: Thank you for your valuable comment on ensuring the reproducibility of our method. We have carefully addressed each of the key elements highlighted, with detailed specifications and explicit formulations supplemented throughout the revised manuscript:

For a) We have mathematically formalized this mechanism in Eq (14), Eq (15), and Eq. (16). Specifically, Eq. (14) clearly defines the specific normalization strategies for the query and key components, and we have provided thorough, step-by-step explanations in the accompanying paragraphs to clarify how the gating weights in Eq (16) are computed.

For b) The formulations for appearance uncertainty and motion uncertainty estimation are explicitly presented in Eq (5) and Eq (9), respectively. We have added detailed definitions of all variables and symbols in these equations, ensuring the mathematical formulation of each uncertainty module is fully specified—no conceptual ambiguity remains, and the implementation logic is clearly elaborated.

For c) The training loss functions corresponding to the uncertainty estimation modules are detailed in Eq (6) and Eq (18). Additionally, the weighting schemes that integrate these uncertainty values into the overall framework are explicitly defined in Eq (4) and Eq (10), with clear descriptions of how these weights are applied during training to optimize the model.

For d) We have supplemented comprehensive Kalman filter formulations for motion uncertainty estimation in Eq (8) and Eq (9), along with clear explanations of all symbols and variables in Section 3.3.2. Furthermore, we have provided the exact definitions of the state transition matrix, measurement matrix, process noise covariance, observation noise covariance, and initial error covariance matrix, along with their specific experimental values to fully support the reproducibility of the motion uncertainty estimation module.

In addition, the pseudocode in Section 3 has been updated accordingly to incorporate all the above specifications, ensuring every key step of the proposed method is fully detailed and can be implemented.

3) Algorithmic clarity and symbol definitions

All variables and symbols appearing in the equations should be explicitly defined. The pseudocode and equations should be sufficiently detailed to allow independent implementation.

Response: Thank you for your valuable comment. We have revised the entire manuscript to explicitly define every variable, symbol, and subscript in all equations, with clear descriptions immediately following each formula in Section 3.

In addition, we have further expanded the pseudocode and mathematical formulations with step-by-step implementation details, including parameter settings, matrix dimensions, iterative procedures, and computation flows in Algorithm 1 and Section 4 model and train details. All critical components are described with sufficient specificity to enable independent re-implementation of the proposed method in the manuscript.

4) Experimental clarity and evaluation reporting

The evaluation section would benefit from clearer reporting of experimental settings, including hyperparameters, memory bank configuration, and training schedules. Where possible, provide additional breakdown of performance (e.g., per-category or scenario-level results) to support the claims regarding improved robustness in challenging scenarios such as occlusion or crowded scenes.

Response: Thank you for your valuable comment. We have detailed a dedicated table that explicitly summarizes all key hyperparameters, including the number of memory banks, learning rate schedule, batch size, optimizer settings, and other related configurations in Table 1 and explained the train details part in Section 4.1.2.

About providing the additional performance, we sincerely appreciate the constructive suggestion on providing more detailed breakdowns to further verify the advantages of UAMP in crowded and occluded scenarios. However, we would like to clarify that the semi-supervised VOS is inherently a class-agnostic task, where given an initial mask, bounding box, or interactive clicks in the first frame, the model only needs to separate the specified target from the background, without relying on semantic category labels. Secondly, none of the compared SOTA methods (Cutie-bas, LiVOS, SAM2.1, SAM2.1Long) reports per-class or per-category J&F metrics in their original papers or official implementations. The official evaluation protocols of the benchmark datasets (e.g., LVOS) also do not provide pre-defined class/category splits for computing such fine-grained per-class statistics. Manually defining arbitrary categories and computing breakdowns alone would lead to incomparable and non-reproducible results against existing SOTAs.

To substantiate our claim that UAMP performs favorably in crowded and occluded scenarios, we evaluated the proposed method on six different datasets, including MOSE, SA-V, LVOSv1, LVOSv2, DAVIS and YouTube-VOS, in which the LVOS is designed for long-term video object segmentation, MOSE, the complex real-life dataset, contains the crowded and occluded scenarios, and SA-V has small, occluded, and reappearing objects in the video sequences. Following the standard evaluation protocol widely accepted in the VOS field, we report overall J&F scores in line with representative recent works. To better substantiate our claim that UAMP performs favorably in crowded and occluded scenarios, we have done the qualitative and quantitative analysis on challenging sequences, highlighting the advantages of our method under complex conditions in Section 4.2.

5) Data availability statement

One reviewer indicated concerns regarding the availability of underlying data. Please ensure that all data required to reproduce the reported findings are publicly available or clearly described in accordance with PLOS ONE’s data availability policy.

Response: Thank you for your reminder regarding the data availability policy. We fully comply with the requirements of PLOS ONE and confirm that all datasets used in this study are publicly accessible standard benchmarks, including MOSE, SA-V, LVOS, DAVIS, and YouTube-VOS, which are widely adopted in the video object segmentation community and freely available to the research community, and download sources are provided in the data availability statement.

We mixed 1% SA-V dataset, 2% MOSE dataset and 1% LVOS dataset for the training dataset, specifically:

1) The SA-V dataset is composed of the videos from sav_000001 to sav_000099 of the sav_000.tar,

2�The MOSE dataset is composed of the videos 10552818 to

Recommended improvements:

1) Improve the structure and clarity of section introductions so that each section and subsection begins with a brief explanation of its purpose.

Response: Thanks for this valuable comment. In the revised version, we have added concise introductory paragraphs at the beginning of the main section to briefly outline the purpose, content, or the organization of the following subsections. These short introductions help guide readers and improve the logical flow of the manuscript in Section 3. We have applied this revision across all relevant sections and subsections to ensure a formal and professional writing style.

2) Ensure consistent terminology throughout the manuscript (e.g., “uncertainty”, “change probability”, “memory bank”, etc.).

Response: Thanks for this valuable comment. We have standardized the terminology throughout the manuscript to maintain consistency. We have unified the use of “uncertainty” instead of inconsistent mixing with “change probability” and we have maintained a clear and consistent distinction between memory bank (referring to the entire memory module and storage structure) and memory (referring to the stored features and information within the bank) to remain consistent with SAM2.

3) Provide a brief discussion of computational overhead and inference latency compared with the baseline SAM2 framework.

Response: Thanks for this valuable comment. We quantitatively compare the proposed UAMP framework with the baseline SAM2 model in terms of memory footprint and inference latency, and analyze the computational complexity and scalability of our method. The corresponding runtime analysis have been added in Section 4.2.2 to clearly demonstrate the efficiency and practicality of UAMP while maintaining performance improvement, which is detailed in Table 9.

4) Review the suggested references provided by the reviewers and cite them where appropriate.

Response: We appreciate the suggested references for potential citation. We have carefully reviewed the research focus of each recommended work:

Paper a) centers on multi-modal feature fusion within a Transformer architecture for integrating visual and audio information to advance dense video captioning. Paper b) explores the design and technological roadmap for developing zero‑emission electric vessels. Paper c) aims at investigating the effect of cigarette smoke exposure on the systemic circulation and local airway 5-HT levels as well as MAO-mediated oxidative pathway using a cigarette smoke-exposed rat model. Paper d) presents device-level simulation and analysis of In₀.₅₃Ga₀.₄₇As implant‑free quantum‑well structures in microelectronics.

Our work targets video object segmentation (VOS), a core task in video understanding that aims to achieve pixel‑level foreground object separation and consistent tracking across consecutive video frames. Given that the above references belong to different research domains, including video captioning, marine engineering, biomedicine, and microelectronic device design, they are not thematically or methodologically aligned with our focus on video object segmentation. Accordingly, we respectfully consider these references less relevant to the present study. We will continue to enrich our literature review by incorporating more state‑of‑the‑art and representative studies dedicated to video object segmentation to better contextualize our contributions for future work.

Reviewer #4

1). The pseudocode in Section 3 is a good addition, but some steps remain underspecified. For example:

a. The gating mechanism for balancing long-term and short-term memory is described qualitatively, but the exact mathematical formulation (activation function, normalization strategy) is missing.

b. The uncertainty estimation modules (appearance vs. motion) are introduced, but the training loss functions and weighting schemes are not fully detailed.

Response: Thanks for your constructive and valuable comments. We have further refined the relevant descriptions and mathematical formulations to clarify the underspecified details in the pseudocode.

Specifically, we have updated the uncertainty estimation modules in Sections 3.3.1 and 3.3.2. The formulations for appearance uncertainty and motion uncertainty estimation are clearly presented in Eq (5) and Eq (9), the corresponding training loss functions are detailed in Eq (6) and Eq (18), where motion uncertainty is modeled using a Kalman filter, which has an analytical solution, and the weighting schemes are defined in Eq (4) and Eq (10). In addition, the gating mechanism used to balance long‑term and short‑term memory has been mathematically formalized in Eq (14), Eq (15) and Eq (16), including the specific activation functions and normalization strategies about query and key, with thorough explanations provided in the paragraphs. The pseudocode in Section 3 has also been updated accordingly to ensure all key steps are fully specified and reproducible.

2). The author(s) introduces additional modules (uncertainty prediction, dual memory mechanisms), but there is no discussion of computational overhead. How does UAMP scale in terms of memory footprint and inference latency compared to SAM2?

Response: We appreciate the reviewer for this comment. In the revised manuscript, we have supplemented a detailed discussion on the computational overhead introduced by the proposed uncertainty prediction and memory management modules. We quantitatively compare the proposed UAMP framework with the baseline SAM2 model in terms of memory footprint and inference latency, and analyze the computational complexity and scalability of our method. The corresponding runtime analysis have been added in Section 4.2.2 to clearly demonstrate the efficiency and practicality of UAMP while maintaining performance improvement, which is detailed in Table 9.

3). Although recent SOTA methods (Cutie-bas, LiVOS, SAM2.1, SAM2.1Long) are included, the evaluation could be expanded: Report per-class or per-category breakdowns of J&F improvements to substantiate claims that UAMP excels in crowded or occluded scenarios.

Response: We sincerely appreciate the constructive suggestion on providing more detailed breakdowns to further verify the advantages of UAMP in crowded and occluded scenarios.

However, we would like to clarify that the semi-supervised VOS is inherently a class-agnostic task, where given an initial mask, bounding box, or interactive clicks in the first frame, the model only needs to separate the specified target from the background, without relying on semantic category labels. Secondly, none of the compared SOTA methods (Cutie-bas, LiVOS, SAM2.1, SAM2.1Long) reports per-class or per-category J&F metrics in their original papers or official implementations. The official evaluation protocols of the benchmark datasets (e.g., LVOS) also do not provide pre-defined class/category splits for computing such fine-grained per-class statistics. Manually defining arbitrary categories and computing breakdowns alone would lead to incomparable and non-reproducible results against existing SOTAs.

To substantiate our claim that UAMP performs favorably in crowded and occluded scenarios, we evaluated the proposed method on six different datasets, including MOSE, SA-V, LVOSv1, LVOSv2, DAVIS and YouTube-VOS, in which the LVOS is designed for long-term video object segmentation, MOSE, the complex real-life dataset, contains the crowded and occluded scenarios, and SA-V has small, occluded, and reappearing objects in the video sequences. Following the standard evaluation protocol widely accepted in the VOS field, we report overall J&F scores in line with representative recent works. To better substantiate our claim that UAMP performs favorably in crowded and occluded scenarios, we have done the qualitative and quantitative analysis on challenging sequences, highlighting the advantages of our method under complex conditions in Section 4.2.

4). The manuscript alternat

Attachments

Attachment

Submitted filename: Response_to_Reviewers_auresp_2.docx

Dear Editor,

We appreciate you and the reviewers for your careful evaluation and constructive suggestions on our manuscript. We have comprehensively revised the manuscript point by point in accordance with all comments raised by the reviewers. All revisions are highlighted in the revised manuscript with track changes. A detailed point-by-point response to each comment is provided below.

Required revisions:

1) Clarification and formal definition of uncertainty modules

The manuscript should clearly define the appearance and motion uncertainty components and describe how these quantities are supervised or learned during training. The current description remains partly conceptual and does not fully specify the mathematical formulation.

Response: Thank you for your careful and constructive comment. We have further strengthened the mathematical definition and learning mechanism of appearance uncertainty and motion uncertainty to avoid conceptual ambiguity.

Specifically, we have clearly defined the formulation of appearance uncertainty in Eq (5) and its corresponding supervised loss in Eq (6), which explicitly guides the learning of appearance reliability during training. For motion uncertainty, we have provided a precise mathematical formulation in Eq (8-9). In the revised manuscript, we have also added detailed explanations for each component, clarifying how these uncertainty values are modeled, optimized, and integrated into the overall training pipeline. All components are now fully specified with explicit mathematical expressions, ensuring the description is concrete and reproducible.

2) Reproducibility of the proposed framework

Several key elements of the method require clearer specification to ensure reproducibility, including:

a) the gating mechanism used for balancing long-term and short-term memory

b) the exact formulation of the uncertainty estimation modules

c) the loss functions and weighting schemes used during training

d) the full specification of the Kalman filter used for motion scoring

Response: Thank you for your valuable comment on ensuring the reproducibility of our method. We have carefully addressed each of the key elements highlighted, with detailed specifications and explicit formulations supplemented throughout the revised manuscript:

For a) We have mathematically formalized this mechanism in Eq (14), Eq (15), and Eq. (16). Specifically, Eq. (14) clearly defines the specific normalization strategies for the query and key components, and we have provided thorough, step-by-step explanations in the accompanying paragraphs to clarify how the gating weights in Eq (16) are computed.

For b) The formulations for appearance uncertainty and motion uncertainty estimation are explicitly presented in Eq (5) and Eq (9), respectively. We have added detailed definitions of all variables and symbols in these equations, ensuring the mathematical formulation of each uncertainty module is fully specified—no conceptual ambiguity remains, and the implementation logic is clearly elaborated.

For c) The training loss functions corresponding to the uncertainty estimation modules are detailed in Eq (6) and Eq (18). Additionally, the weighting schemes that integrate these uncertainty values into the overall framework are explicitly defined in Eq (4) and Eq (10), with clear descriptions of how these weights are applied during training to optimize the model.

For d) We have supplemented comprehensive Kalman filter formulations for motion uncertainty estimation in Eq (8) and Eq (9), along with clear explanations of all symbols and variables in Section 3.3.2. Furthermore, we have provided the exact definitions of the state transition matrix, measurement matrix, process noise covariance, observation noise covariance, and initial error covariance matrix, along with their specific experimental values to fully support the reproducibility of the motion uncertainty estimation module.

In addition, the pseudocode in Section 3 has been updated accordingly to incorporate all the above specifications, ensuring every key step of the proposed method is fully detailed and can be implemented.

3) Algorithmic clarity and symbol definitions

All variables and symbols appearing in the equations should be explicitly defined. The pseudocode and equations should be sufficiently detailed to allow independent implementation.

Response: Thank you for your valuable comment. We have revised the entire manuscript to explicitly define every variable, symbol, and subscript in all equations, with clear descriptions immediately following each formula in Section 3.

In addition, we have further expanded the pseudocode and mathematical formulations with step-by-step implementation details, including parameter settings, matrix dimensions, iterative procedures, and computation flows in Algorithm 1 and Section 4 model and train details. All critical components are described with sufficient specificity to enable independent re-implementation of the proposed method in the manuscript.

4) Experimental clarity and evaluation reporting

The evaluation section would benefit from clearer reporting of experimental settings, including hyperparameters, memory bank configuration, and training schedules. Where possible, provide additional breakdown of performance (e.g., per-category or scenario-level results) to support the claims regarding improved robustness in challenging scenarios such as occlusion or crowded scenes.

Response: Thank you for your valuable comment. We have detailed a dedicated table that explicitly summarizes all key hyperparameters, including the number of memory banks, learning rate schedule, batch size, optimizer settings, and other related configurations in Table 1 and explained the train details part in Section 4.1.2.

About providing the additional performance, we sincerely appreciate the constructive suggestion on providing more detailed breakdowns to further verify the advantages of UAMP in crowded and occluded scenarios. However, we would like to clarify that the semi-supervised VOS is inherently a class-agnostic task, where given an initial mask, bounding box, or interactive clicks in the first frame, the model only needs to separate the specified target from the background, without relying on semantic category labels. Secondly, none of the compared SOTA methods (Cutie-bas, LiVOS, SAM2.1, SAM2.1Long) reports per-class or per-category J&F metrics in their original papers or official implementations. The official evaluation protocols of the benchmark datasets (e.g., LVOS) also do not provide pre-defined class/category splits for computing such fine-grained per-class statistics. Manually defining arbitrary categories and computing breakdowns alone would lead to incomparable and non-reproducible results against existing SOTAs.

To substantiate our claim that UAMP performs favorably in crowded and occluded scenarios, we evaluated the proposed method on six different datasets, including MOSE, SA-V, LVOSv1, LVOSv2, DAVIS and YouTube-VOS, in which the LVOS is designed for long-term video object segmentation, MOSE, the complex real-life dataset, contains the crowded and occluded scenarios, and SA-V has small, occluded, and reappearing objects in the video sequences. Following the standard evaluation protocol widely accepted in the VOS field, we report overall J&F scores in line with representative recent works. To better substantiate our claim that UAMP performs favorably in crowded and occluded scenarios, we have done the qualitative and quantitative analysis on challenging sequences, highlighting the advantages of our method under complex conditions in Section 4.2.

5) Data availability statement

One reviewer indicated concerns regarding the availability of underlying data. Please ensure that all data required to reproduce the reported findings are publicly available or clearly described in accordance with PLOS ONE’s data availability policy.

Response: Thank you for your reminder regarding the data availability policy. We fully comply with the requirements of PLOS ONE and confirm that all datasets used in this study are publicly accessible standard benchmarks, including MOSE, SA-V, LVOS, DAVIS, and YouTube-VOS, which are widely adopted in the video object segmentation community and freely available to the research community. And for training our model,

We mixed about 1% SA-V dataset, 20% MOSE dataset and 20% LVOS dataset for the training dataset, specifically:

1) The SA-V dataset is composed of the middle 20 videos from sav_000491 to sav_000510 of the sav_trian.tar file, which can be download form the link https://ai.meta.com/datasets/segment-anything-video-downloads/.

2�The first 300 videos are extracted from the MOSE dataset according to the meta_train.json file contained in the train.tar.gz dataset, which can be freely downloaded from the link https://huggingface.co/datasets/FudanCVL/MOSE/tree/main.

3) The last 50 videos are chosen from the LVOS v1 dataset and last 150 videos from the LVOS v2 dataset according to the meta.json contained in the train.zp file, which can be freely downloaded from the link https://lingyihongfd.github.io/lvos.github.io/dataset.html.

And we have added the training data details in the section 4.1.2, and the updated the data availability statement in the revision.

Recommended improvements:

1) Improve the structure and clarity of section introductions so that each section and subsection begins with a brief explanation of its purpose.

Response: Thanks for this valuable comment. In the revised version, we have added concise introductory paragraphs at the beginning of the main section to briefly outline the purpose, content, or the organization of the following subsections. These short introductions help guide readers and improve the logical flow of the manuscript in Section 3. We have applied this revision across all relevant sections and subsections to ensure a formal and professional writing style.

2) Ensure consistent terminology throughout the manuscript (e.g., “uncertainty”, “change probability”, “memory bank”, etc.).

Response: Thanks for this valuable comment. We have standardized the terminology throughout the manuscript to maintain consistency. We have unified the use of “uncertainty” instead of inconsistent mixing with “change probability” and we have maintained a clear and consistent distinction between memory bank (referring to the entire memory module and storage structure) and memory (referring to the stored features and information within the bank) to remain consistent with SAM2.

3) Provide a brief discussion of computational overhead and inference latency compared with the baseline SAM2 framework.

Response: Thanks for this valuable comment. We quantitatively compare the proposed UAMP framework with the baseline SAM2 model in terms of memory footprint and inference latency, and analyze the computational complexity and scalability of our method. The corresponding runtime analysis have been added in Section 4.2.2 to clearly demonstrate the efficiency and practicality of UAMP while maintaining performance improvement, which is detailed in Table 9.

4) Review the suggested references provided by the reviewers and cite them where appropriate.

Response: We appreciate the suggested references for potential citation. We have carefully reviewed the research focus of each recommended work:

Paper a) centers on multi-modal feature fusion within a Transformer architecture for integrating visual and audio information to advance dense video captioning. Paper b) explores the design and technological roadmap for developing zero‑emission electric vessels. Paper c) aims at investigating the effect of cigarette smoke exposure on the systemic circulation and local airway 5-HT levels as well as MAO-mediated oxidative pathway using a cigarette smoke-exposed rat model. Paper d) presents device-level simulation and analysis of In₀.₅₃Ga₀.₄₇As implant‑free quantum‑well structures in microelectronics.

Our work targets video object segmentation (VOS), a core task in video understanding that aims to achieve pixel‑level foreground object separation and consistent tracking across consecutive video frames. Given that the above references belong to different research domains, including video captioning, marine engineering, biomedicine, and microelectronic device design, they are not thematically or methodologically aligned with our focus on video object segmentation. Accordingly, we respectfully consider these references less relevant to the present study. We will continue to enrich our literature review by incorporating more state‑of‑the‑art and representative studies dedicated to video object segmentation to better contextualize our contributions for future work.

-------------------------------------------------------------------------------------------------------

Reviewer #4

1). The pseudocode in Section 3 is a good addition, but some steps remain underspecified. For example:

a. The gating mechanism for balancing long-term and short-term memory is described qualitatively, but the exact mathematical formulation (activation function, normalization strategy) is missing.

b. The uncertainty estimation modules (appearance vs. motion) are introduced, but the training loss functions and weighting schemes are not fully detailed.

Response: Thanks for your constructive and valuable comments. We have further refined the relevant descriptions and mathematical formulations to clarify the underspecified details in the pseudocode, specifically:

For a, the gating mechanism used to balance long‑term and short‑term memory has been mathematically formalized in Eq (14), Eq (15) and Eq (16), where Eq. (14) clearly defines the specific normalization strategies for the query and key components, and we have provided thorough, step-by-step explanations in the accompanying paragraphs to clarify how the gating weights in Eq (16) are computed.

For b, we have updated the uncertainty estimation modules in Sections 3.3.1 and 3.3.2. The formulations for appearance uncertainty and motion uncertainty estimation are clearly presented in Eq (5) and Eq (9), the corresponding training loss functions are detailed in Eq (6) and Eq (18), where motion uncertainty is modeled using a Kalman filter, which has an analytical solution, and the weighting schemes are defined in Eq (4) and Eq (10). The pseudocode in Section 3 has also been updated accordingly in the revision, to ensure all key steps are fully specified and reproducible.

2). The author(s) introduces additional modules (uncertainty prediction, dual memory mechanisms), but there is no discussion of computational overhead. How does UAMP scale in terms of memory footprint and inference latency compared to SAM2?

Response: We appreciate the reviewer for this comment. In the revised manuscript, we have supplemented a detailed discussion on the computational overhead introduced by the proposed uncertainty prediction and memory management modules. We quantitatively compare the proposed UAMP framework with the baseline SAM2 model in terms of memory footprint and inference latency, and analyze the computational complexity and scalability of our method. The corresponding runtime analysis have been added in Section 4.2.2 to clearly demonstrate the efficiency and practicality of UAMP while maintaining performance improvement, which is detailed in Table 9.

3). Although recent SOTA methods (Cutie-bas, LiVOS, SAM2.1, SAM2.1Long) are included, the evaluation could be expanded: Report per-class or per-category breakdowns of J&F improvements to substantiate claims that UAMP excels in crowded or occluded scenarios.

Response: We sincerely appreciate the constructive suggestion on providing more detailed breakdowns to further verify the advantages of UAMP in crowded and occluded scenarios.

However, we would like to clarify that the semi-supervised VOS is inherently a class-agnostic task, where given an initial mask, bounding box, or interactive clicks in the first frame, the model only needs to separate the specified target from the background, without relying on semantic category labels. Secondly, none of the compared SOTA methods (Cutie-bas, LiVOS, SAM2.1, SAM2.1Long) reports per-class or per-category J&F me

Attachments

Attachment

Submitted filename: Response_to_Reviewers_auresp_3.docx