CamSAM2: Segment Anything Accurately in Camouflaged Videos

We sincerely thank you for the positive feedback and thoughtful suggestions. We are glad that you found our research motivation to be clear and our method design to be easy to follow. We also appreciate your recognition of the architectural clarity and the integration of the IOF and EOF modules based on the SAM2 framework. Your comments encourage us to further refine and clarify the technical aspects of our work, and we address your insightful questions and concerns below.

---

W1: OPG details clarification

Thanks for your thoughtful questions.

1. Regarding the number of prototypes ($k$), we have investigated its impact in Table 7 in the main paper. The results show that both fewer or higher numbers of prototypes will reduce the performance due to under-representation or redundancy, but are still effective in abstracting prototypes.
2. Regarding the consistency of prototypes across time, these prototypes remain consistent in a semantic sense, as they are all abstracted from the previously predicted camouflaged regions. Since these regions represent the same object over time, the resulting prototypes serve as temporally coherent and semantically aligned for guiding the segmentation in subsequent frames.

To improve interpretability, we calculate cosine similarity maps between each prototype of previous frames and the feature map of the current frame. The visualizations reveal that different prototypes consistently activate on semantically meaningful object regions. For example, the prototype shows strong activations (high similarity) precisely on the camouflaged object, while suppressing the background. This suggests that the prototypes successfully abstract object-aware semantics and are consistent across time. We will present detailed visualizations and analysis in the final version of the paper to further support interpretability.

---

W2: Real-time capability concern

Thanks for raising this practical concern. We profile the runtime overhead of CamSAM2 on a 106-frame video and report the average per-frame latency in Table 1. CamSAM2 adds only 6 ms latency per frame on top of SAM2, resulting in a total per-frame time of 89.7 ms. This demonstrates the model's practicality in real-time applications such as wildlife monitoring, search and rescue.

Table 1. Average runtime per frame and overhead breakdown of CamSAM2 compared to original SAM2.

---

W3: Structural issue with Section 3

Thanks for your suggestion. We agree that Sections 3.6 and 3.7 are more appropriately arranged under the Experiments section. We will move them into Section 4 in the final version to improve structural consistency.

---

W4: Publication year in Table

Thanks for your advice. We will add the publication year in the final version.

---

Q1: More VOS models comparisons

Thanks for your suggestion. We compare CamSAM2 with two recent and representative SOTA video object segmentation (VOS) models: Cutie [1] and XMem++ [2], both widely used in VOS tasks. For fairness, we evaluate all models using the ground truth of the first frame as the mask prompt on the MoCA-Mask. As shown in Table 2, CamSAM2 achieves 70.5 mIoU, surpassing all baselines with comparable model sizes. This demonstrates the strong performance of CamSAM2 over both VOS-specific models and its SAM2-based baseline.

Table 2. Comparison with VOS models.

---

Q2: Highlighting the roles of components

We appreciate your suggestion. In CamSAM2, the decamouflaged token and OPG module are the core components that enable our method to adapt SAM2 for the VCOS task. The decamouflaged token allows the model to learn camouflaged features for camouflaged objects and is used to generate the decamouflaged mask without modifying SAM2's architecture, while OPG abstracts object-aware prototypes from previous frames to guide segmentation across time. The IOF and EOF modules play supporting roles, further enhancing spatial and temporal feature representation. To make this distinction clearer, we will revise the final version to more explicitly highlight the relative importance and motivation behind each module.

---

Limitation

Thanks for raising this important point. We agree that both computational complexity and inference speed are critical considerations. As mentioned in W2, inference speed has already been addressed. Besides, we profile the computational complexity in Table 3. CamSAM2 introduces an additional 14.4 GFLOPs, which is independent of different backbone architectures. Based on the Hiera-T backbone, the total FLOPs increase from 137.2 GFLOPs to 151.5 GFLOPs, which corresponds to approximately a 10% increase. This added cost results in a significant performance gain of 9.8 mIoU with point prompts.
Moreover, when using the Hiera-L backbone, which image encoder requires 810 GFLOPs, the same 14.4 GFLOPs overhead accounts for only a 1.7% increase in total FLOPs. Even with such a small relative increase, CamSAM2 still achieves a notable 7.5 mIoU gain. This highlights the efficiency of our design: the added modules contribute strong segmentation improvements at minimal computational cost.

Table 3. FLOPs comparison between SAM2 and CamSAM2.

---

References
[1] Putting the object back into video object segmentation.
[2] Xmem++: Production-level video segmentation from few annotated frames.

We sincerely thank you for the detailed and constructive feedback. We greatly appreciate your recognition of the clarity of our architectural design and the comprehensive comparisons provided. Now we address each of your concerns below.

---

W1: Usage of prompts on VCOS

Thank you for raising this important point. We would like to clarify that our work specifically addresses the task of VCOS, which has distinct goals from COD. COD typically focuses on detecting camouflaged objects in static images without any prior cues, whereas VCOS requires accurate and temporally consistent segmentation of camouflaged objects across entire video sequences. Although our method uses prompts, this design choice is aligned with the nature of SAM and SAM2 as promptable segmentation models. In fact, recent works such as COMPrompter[1], DSAM[2], and SAM-PM[3] also incorporate box or mask prompts to adapt SAM to the COD or VCOS task. Thus, using prompts is a common and well-accepted practice in recent work and does not contradict the objectives of the task.

What's more, these two paradigms, COD for detection and VCOS for segmentation, can be effectively combined. For example, one could apply a COD model to the first frame of a video to generate a bounding box, which then serves as a box prompt to CamSAM2 for video segmentation. This integration not only makes CamSAM2 compatible with existing COD methods and fully automatic pipelines but also highlights the practical value and significance of our work in advancing VCOS. In real-world wildlife monitoring, search and rescue scenarios, a simple user interaction (e.g. clicking on the target object) can trigger CamSAM2 to segment and track it across time.

---

Q1: Loss design clarification

Thank you for this insightful question. While the original SAM2 parameters are frozen during training, we still apply supervision to its output $\mathbf{R}_i$ for the following reasons: Although SAM2's weights are fixed, the introduction of our learnable decamouflaged token changes the self-attention dynamics in the mask decoder. This token is concatenated with the original output tokens and jointly participates in self-attention and cross-attention layers (see Fig. 2(a) in the main paper). As a result, the outputs $\mathbf{R}_i$ are no longer identical to the original SAM2 outputs. Supervising them ensures that SAM2's outputs remain aligned with the segmentation intent, rather than being perturbed by the new token interactions. As shown in the Supple. Materials Table 4, we report the mIoU scores evaluated with the mask predicted by the SAM2 mask token and the decamouflaged token, the results also demonstrate the effectiveness of our design.

---

Q2: Usage of prompts on VCOS

Thanks for your question. Please refer to W1 in detail.

---

Limitation: On prototype potential overfitting

Thank you for raising this insightful concern. Our OPG and EOF modules are carefully designed to mitigate this risk. First, prototypes are dynamically generated from the predicted camouflaged regions using FPS and 1-iter $k$-means. This ensures that the prototypes are abstracted from the entire object masks, and are continually updated across frames and adapt to object appearance changes. Second, the EOF module uses a cross-attention mechanism, where the current frame's features query the prototype memory. This structure naturally suppresses outdated or irrelevant prototypes: if a prototype is inconsistent or suboptimal (e.g., due to occlusion or missing), it receives lower attention and contributes minimally to the output. This mechanism provides robustness and avoids overfitting to specific regions. Our Fig. 3 in the main paper further demonstrates this, showing that the segmentation results remain stable across time and do not overfit to a specific local region.

---

References
[1] COMPrompter: Reconceptualized Segment Anything Model with Multiprompt Network for Camouflaged Object Detection
[2] Exploring Deeper! Segment Anything Model with Depth Perception for Camouflaged Object Detection
[3] SAM-PM: Enhancing Video Camouflaged Object Detection using Spatio-Temporal Attention

We are glad to hear that our paper is well-presented, with clear architectural diagrams, coherent mathematical formulations, and strong empirical validation across multiple datasets. Below, we address your concerns point by point.

W1: IOF novelty clarification

We appreciate your comment. The IOF module is designed to address a key limitation in SAM2, where the early-stage high-resolution features are not fully utilized during the segmentation process. These features contain fine-grained texture and boundary details are especially crucial for VCOS. To exploit this, IOF fuses early-layer high-resolution features with memory-conditioned semantic features, thereby enhancing the model's ability to perceive subtle camouflage cues. While IOF may share structural similarities with general fusion mechanisms (e.g., FPN), its motivation, integration strategy, and application context are specifically tailored for VCOS. To the best of our knowledge, this is the first work to incorporate early-layer high-resolution features into SAM2 for VCOS. Our ablation study (Table 5 in the main paper) shows that adding IOF brings a non-trivial performance gain (e.g., +0.5 mIoU), highlighting both its necessity and effectiveness.

---

W2: Explanation of design choices

Thanks for pointing this out. We clarify these design decisions as follows:

1. Decamouflaged token: The decamouflaged token is designed with a shape of 1 × 256. As illustrated in Fig. 2 in the main paper, this decamouflaged token extends SAM2's token structure, aligning with the default embedding dimension of output tokens in SAM2. This ensures compatibility with the existing self-attention structure without modifying SAM2's architecture.
2. Number of prototypes in OPG: The number of prototypes in OPG is fixed (set to k = 5) across all datasets. We initially tuned this value on MoCA-Mask and found it to generalize well, as shown in our ablation study in the main paper Table 7. Our goal is not precise clustering, but to abstract object-aware prototypes to guide cross-attention in EOF. A fixed k also simplifies inference and avoids introducing dataset-specific hyperparameters.

We will clearly add these explanations in the final version. In addition, we will publicly release our code to support reproducibility and transparency.

---

W3: On the placement of mask comparisons

Thanks for the suggestion. In the final version, we will consider moving this analysis into the main paper to better showcase how our design benefits segmentation accuracy.

---

W4: Minor typo

Thanks for pointing out our typo. We will fix it and carefully proofread the entire manuscript to eliminate other minor errors in the final version.

---

Q1: Some suggestions

Thank you for the suggestion. We agree that distinguishing major and minor components can help clarify our contributions. We will emphasize this distinction more clearly in the revised version.

---

Q2: Explanation of design choices

Thanks for your question. Please refer to W2 in detail.

Thank you for your valuable suggestions and for recognizing the strengths of our work, including the strong empirical performance and the generalization ability of CamSAM2 across diverse VCOS scenarios. Below, we respond in detail to your concerns.

---

W1. Novelty clarification

CamSAM2 introduces a cohesive and task-specific redesign tailored for the challenging VCOS task. Concretely, CamSAM2 addresses two key limitations of SAM2 in the VCOS scenarios: (i) SAM2 is optimized for natural scene segmentation and struggles with the fine-grained, low-contrast textures characteristic of camouflaged objects; (ii) SAM2's memory module stores only coarse, low-resolution features, which impairs its ability to maintain temporal consistency across time.

To tackle these, we introduce a set of interdependent modules, which are co-designed rather than independently stacked. This integration in a frozen SAM2 pipeline is novel and non-trivial. To the best of our knowledge, CamSAM2 is the first work to successfully and efficiently adapt SAM2 to the VCOS task, achieving both architectural innovation and practical performance gains.

---

W2. Efficiency concerns of OPG (FPS and k-means)

Thank you for highlighting the potential runtime overhead introduced by our proposed modules, especially the OPG. To address this, we profile CamSAM2 with hiera-T as the backbone on a 106-frame camouflaged video from the MoCA-Mask test set, and report the average per-frame latency in Table 1. For comparison, we also include the original SAM2 per-frame runtime.

Table 1. Average runtime per frame and overhead breakdown of CamSAM2 compared to original SAM2.

These results confirm that CamSAM2 introduces only 6 ms (7.2%) additional latency per frame compared to SAM2. As you correctly noted, the OPG module (comprising both FPS and 1-iter k-means clustering) is indeed the dominant component of overhead. Apart from OPG, all other components, including IOF, EOF, and associated processing, together contribute only approximately 1.4 ms (1.7%) on average per frame. These results demonstrate that CamSAM2 remains efficient and is suitable for practical deployment, even in time-sensitive VCOS applications.

---

W3. Insufficient analysis and ablations

1. The number of the decamouflaged token

Thanks for your suggestion. We conduct an ablation study by increasing the number of tokens to 3 and observed a comparable result in mIoU, as shown in Table 2. Adding more decamouflaged tokens does not structurally improve the representation quality but may introduce redundancy and computational cost.

Table 2. Impact of using different numbers of decamouflaged tokens.

---

2. Sampling and clustering strategies

To assess the robustness and generality of OPG, we experiment on different sampling and clustering strategies: (i) Farthest Point Sampling (FPS) vs. Average Sampling for selecting initial cluster centers, and (ii) k-means vs. Gaussian Mixture Model (GMM) as the clustering algorithm.

As shown in Table 3, all combinations yield reasonable results. Compared to average sampling, FPS better preserves spatial diversity among selected features, leading to more representative and discriminative prototypes. These results still confirm that OPG is flexible and not overly sensitive to specific implementation choices, which enhances its applicability.

Table 3. Impact of different sampling and clustering strategies in OPG.

---

3. Prototype visualization and discussion

Thanks for pointing this out. Following your suggestion, we compute the similarity map between each prototype of previous frames and the feature map of the current frame. These maps reflect how each prototype activates spatially over the image.

Our visualizations show that different prototypes consistently activate on semantically meaningful object regions. For example, the prototype shows strong activations (high similarity) precisely on the camouflaged object, while suppressing the background. This suggests that the prototypes successfully abstract object-aware semantics, supporting their effectiveness in guiding the cross-attention in EOF. Although visual examples cannot be included in this rebuttal, we will add these visualizations and related analysis in the final version of the paper.

---

Minor Weaknesses

1. Thanks for your suggestions, we will refine the figures in the final version.

2. We appreciate your attention to this point. To clarify, in Sec. 7 of the Supple. Material, we describe the mode switch between CamSAM2 and SAM2 at inference. When the toggle is off, all parameters introduced by CamSAM2 are deactivated, and the model functions identically to SAM2 without any parameter modifications. We also test on the SA-V test set and get the same J&F result. This design ensures full architectural compatibility and serves as the basis for our claim about retaining the original segmentation capabilities on natural images and videos.

3. Thank you for pointing this out. The camouflaged objects in our benchmark are indeed highly challenging to perceive due to their nature. The ground-truth masks in Fig. 3 in the main paper are sourced directly from the MoCA-Mask dataset, which provides high-quality human-annotated segmentation masks. While the visualization may appear unclear in certain cases because the objects are highly camouflaged, the annotations are accurate and widely accepted in the community.

Dear Reviewer,

Thank you very much for your thoughtful feedback on our submission. We have carefully addressed your comments in our response.

If you have any additional questions or concerns, we would be more than happy to clarify.

We sincerely appreciate your time and efforts in reviewing our work.