acceptodds
Withdrawn from ICLR 2027

CamMem: Structured Camouflage-State Memory Replay with SAM2 for Camouflaged Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Camouflaged object detection (COD) aims to segment objects concealed by their surroundings, where high foreground–background similarity and indistinct boundaries make accurate perception particularly challenging. Although Segment Anything Model 2 (SAM2) provides strong visual representations and memory mechanisms for video propagation, these mechanisms are not naturally tailored to single-image camouflage perception. In this work, we propose CamMem, a structured camouflage-state memory replay framework that adapts SAM2 to COD through an intra-image memory writing–memory replay–conflict correction process. Specifically, we first introduce Structured Camouflage-State Memory (SCSM), which organizes coarse predictions and visual features into task-specific camouflage states encoding foreground-background, boundary, reliability, and conflict cues. We then develop a Camouflage-State Replay Adapter (CSRA), which combines relation-calibrated state matching, evidence-directed replay, and spatial–channel gated injection to selectively propagate structured camouflage states across SAM2 encoder blocks. Finally, Conflict-Guided Anchor Correction (CGAC) retains the replay-conditioned prediction as the primary hypothesis and uses the initial prediction as a stable anchor, regulating anchor correction through prediction conflict and local-support preference with bounded residual mixing. Experimental results on four COD benchmarks, including CAMO, COD10K, NC4K, and CHAMELEON, show that CamMem achieves substantial performance gains over the SAM2-based baseline and delivers competitive results against state-of-the-art COD methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Camouflaged object detection, Segment Anything Model 2, structured camouflage-state memory, memory replay, conflict-guided correction.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_W2dcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CamMem addresses camouflaged object detection by constructing five overlapping state prototypes from an initial prediction of the same image. It replays these prototypes into SAM2 encoder features and selectively corrects the resulting prediction using the initial output as an anchor. The paper evaluates four camouflage benchmarks, component ablations, paired per-image results, and extensions to salient object detection, defocus blur detection, and forgery localization.

Critical strengths
  1. The method integrates image-specific state aggregation, signed encoder feedback, and bounded correction into a concrete refinement procedure. The correction ablation provides evidence of useful component interaction.
  2. The empirical value extends beyond favorable aggregate scores. The separately trained defocus blur experiment also demonstrates broader applicability under a stated protocol, obtaining good performance on 500 DUT images after training on 604 CUHK images.
Critical weaknesses
  1. The incremental value of the specific state decomposition remains insufficiently isolated from established prediction-conditioned regional context. CamMem adds five overlapping states, relation calibration, signed encoder replay, and anchor correction, which constitute identifiable architectural changes. However, the reduced-state controls do not specify the accompanying changes to state-dependent replay, leaving the contribution of the additional state distinctions entangled with changes to the replay operator.
  2. The paper does not sufficiently establish why the proposed components address camouflage specifically or why they also benefit SOD. SCSM derives its state weights from prediction confidence and boundary variation, while CGAC uses prediction disagreement and local support, but the analyses do not verify that these quantities identify camouflage-related errors. However, Appendix C.1 reports SOD scores without specifying the training and transfer protocol or providing component ablations, preventing a distinction between generic segmentation refinement and transfer of a camouflage-specific mechanism.
Questions for authors
  1. What evidence shows that the "reliable" and "conflict" states identify actual prediction errors associated with weak foreground and background separation, and that replay or anchor correction preferentially reduces those errors?
  2. Does Table 6 evaluate COD-trained weights directly or a separately trained SOD model, and what component-level evidence under that protocol supports the claimed transfer of the memory mechanism? This distinction is necessary to explain whether the opposing appearance characteristics of COD and SOD require different learned behavior within the same architecture.
  3. How are the Unified and FG + BG variants implemented when the reliable and conflict prototypes and their associated relation and directed replay terms are unavailable?
  4. What is the complete training procedure linking the initial and replay branches, including the initial checkpoint’s provenance, prompt inputs, and the prediction outputs receiving the composite loss? They should be clearified in the paper.
  5. Were all baseline results evaluated using consistent metric definitions, given that Table 1 places SAM2-Adapter’s COD10K score of 0.950, originally reported as mean E-measure, under maximum E-measure? Standard evaluation in SOD, COD and related segmentation tasks distinguishes maximum, mean and adaptive variants of E-measure and F-measure, as well as the separate weighted F-measure, so this mismatch raises substantive concerns about the consistency of the other baseline comparisons. When an original paper does not report the required variant, it should be recomputed from the authors' released prediction maps or predictions generated using their official code and trained weights under a common evaluation protocol. An entry should remain unavailable only when neither a compatible published result nor usable prediction or reproduction resources can be obtained, with the reason explicitly stated.
6 comments

We sincerely thank the reviewer for the detailed and constructive feedback. We particularly appreciate your recognition of CamMem's integration of image-specific memory states, evidence-directed replay, and bounded correction, as well as its applicability to other dense prediction tasks. We also acknowledge your concerns regarding the effectiveness of individual memory states, cross-task generalization, implementation details, and evaluation consistency. To address these concerns, we have conducted additional state–error alignment and correction analyses, supplemented the SOD experiments with component-level ablations, and further clarified the memory-composition controls and two-stage training procedure. We have also carefully reviewed the baseline evaluation metrics and are correcting the identified inconsistencies. We provide detailed responses to each concern below and will incorporate these improvements into the revised manuscript.


Q1: What evidence shows that the "reliable" and "conflict" states identify actual prediction errors associated with weak foreground and background separation, and that replay or anchor correction preferentially reduces those errors?

A1: We thank the reviewer for this important question. REL and CONF encode prediction-derived evidence rather than supervised error labels. To examine their correspondence with actual errors and subsequent error recovery, we conducted a state–error alignment analysis and independent model comparison on NC4K .

High-REL and High-CONF comprise pixels whose state responses strictly exceed their respective per-image 80th percentiles, excluding cutoff ties. The responses are obtained from a fixed reference Initial prediction, and the resulting regions remain unchanged across model variants and may overlap. Errors are evaluated at the original ground-truth resolution using a probability threshold of 0.5 without min–max normalization. Recovery denotes the percentage of the same reference Initial error pixels correctly predicted by each model. Ground truth is used only for offline evaluation.

R3-Table 1. State–error alignment and reference-error recovery on NC4K.

RegionInitial Error (%)w/o CGAC Recovery (%)Full Recovery (%)
High-REL17.0084.7985.19
High-CONF20.6881.5682.49

(1) State–error correspondence. High-CONF exhibits a higher Initial error rate than High-REL (20.68% versus 17.00%), consistent with their intended distinction between conflicting and reliable evidence. Importantly, REL reflects reliability estimated from initial perception, not ground-truth correctness; confident but incorrect predictions can therefore occur within High-REL. The results support an observed difference in prediction reliability between these state-associated regions, rather than interpreting either state as an explicit error detector.

(2) Error recovery. CamMem without CGAC recovers 84.79% and 81.56% of reference errors in High-REL and High-CONF, respectively. Full CamMem increases these rates to 85.19% and 82.49%. Both configurations therefore recover a substantial proportion of initial errors within these regions, with modest additional recovery for the full configuration.

These findings provide evidence of error recovery within both state-associated regions. However, the comparison does not isolate the contribution of CGAC or establish preferential correction relative to other regions. We will incorporate these results and clarify their interpretation in the revised manuscript.


Q2: Does Table 6 evaluate COD-trained weights directly or a separately trained SOD model, and what component-level evidence under that protocol supports the claimed transfer of the memory mechanism? This distinction is necessary to explain whether the opposing appearance characteristics of COD and SOD require different learned behavior within the same architecture.

A2: We thank the reviewer for raising this important question. We clarify the training protocol and provide additional component-level ablations to evaluate the effectiveness of CamMem on SOD.

(1) Training protocol. Table 6 evaluates a separately trained SOD model rather than directly applying COD-trained weights to SOD benchmarks. Specifically, CamMem is trained on DUTS-TR for SOD, while the COD model is trained on the training subsets of CAMO and COD10K. This task-specific training protocol is also adopted in recent studies, such as Controllable-LPMoE [1] and ESCNet [2]. Both tasks employ the same CamMem architecture, including SCSM, CSRA, and CGAC, but use independently learned weights. Therefore, Table 6 demonstrates the applicability of the proposed architecture across COD and SOD, rather than zero-shot transfer of COD-trained weights.

(2) Component-level ablation on SOD. To further examine whether the proposed memory mechanism benefits SOD, we conducted additional component-level ablations on DUTS-TE and PASCAL-S under the SOD training protocol. The results are summarized in R3-Table 2.

R3-Table 2: Ablation study of the main CamMem components on SOD benchmarks.

CSRASCSMCGACDUTS-TE M↓Max F↑Sm↑Em↑PASCAL-S M↓Max F↑Sm↑Em↑
———0.0250.9330.9180.9470.0500.8970.8600.905
✓——0.0220.9360.9270.9550.0470.9070.8830.920
✓✓—0.0190.9440.9350.9600.0440.9120.8940.935
✓✓✓0.0180.9470.9390.9620.0420.9160.8990.933

As shown in R3-Table 2, progressively incorporating CSRA, SCSM, and CGAC improves the overall SOD performance. On DUTS-TE, MAE decreases from 0.025 to 0.018, while maximum F-measure and S-measure increase from 0.933 and 0.918 to 0.947 and 0.939, respectively. Similar improvements are observed on PASCAL-S, where MAE decreases from 0.050 to 0.042, and maximum F-measure and S-measure increase from 0.897 and 0.860 to 0.916 and 0.899. Although removing CGAC yields a slightly higher mean E-measure on PASCAL-S (0.935 versus 0.933), the complete model achieves the best MAE, maximum F-measure, and S-measure on both datasets. These results provide component-level evidence that structured memory construction, selective replay, and anchor correction also benefit SOD.

(3) Cross-task applicability. Although COD and SOD exhibit different foreground–background appearance characteristics, CamMem constructs memory states from prediction-derived confidence, boundary, and conflict cues rather than predefined object appearances. The same architecture can therefore learn task-adapted representations and memory interactions through separate training. The additional ablations support the effectiveness of this shared architectural design across both tasks, while not implying direct transferability of COD-trained weights or identical learned behavior between COD and SOD.

We will clarify the training protocol and incorporate these additional ablation results into the revised manuscript.

References

[1] Yanguang Sun, Jiawei Lian, Jian Yang, and Lei Luo. Controllable-LPMoE: Adapting to challenging object segmentation via dynamic local priors from mixture-of-experts. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 22327–22337, 2025.

[2] Sheng Ye, Xin Chen, Yan Zhang, Xianming Lin, and Liujuan Cao. ESCNet: Edge-semantic collaborative network for camouflaged object detection. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 20053–20063, 2025.

Q3: How are the Unified and FG + BG variants implemented when the reliable and conflict prototypes and their associated relation and directed replay terms are unavailable?

A3: We thank the reviewer for pointing out this ambiguity. Both Unified and FG + BG retain the five-slot memory interface. Unavailable state-specific representations are replaced by shared representations rather than physically removed.

(1) Unified. All five slots receive the same globally averaged feature, together with the mean of the five state embeddings. The shared projection and normalization therefore produce identical memory tokens, removing state-specific content and identity information.

(2) FG + BG. The foreground and background slots retain their respective weighted feature aggregates and embeddings. The BD, REL, and CONF slots receive the arithmetic mean of the FG/BG feature aggregates and the mean of their embeddings. Replacement occurs before the shared prototype projection and final normalization.

(3) Relation calibration and directed replay. The CSRA computation, including relation priors and spatial–channel gates, remains enabled. In FG + BG, the identical REL and CONF tokens yield identical keys and matching scores. Their signed evidence difference is therefore zero, making the corresponding directed replay term vanish, even though relation calibration can produce different values for these slots. In Unified, both FG–BG and REL–CONF matching-score differences vanish.

However, these variants do not disable the entire directed replay operation: attention retrieval and the boundary-associated residual term remain active, with the latter using the replacement BD representation. Likewise, the relation priors retain their original slot-indexed structure, although the replaced slots no longer contain the corresponding state-specific evidence.

These experiments are therefore memory-composition controls under a fixed replay architecture, not strict state-deletion ablations or independent tests of relation branches. We will clarify these replacement rules and their downstream implications in the revised manuscript.


Reviewer_hcA5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of segmenting camouflaged objects in single images. It focuses on how to organize and reuse information from a model’s own initial predictions to mitigate perceptual challenges arising from high foreground-background similarity and ambiguous object boundaries. The authors propose CamMem, a framework based on SAM2, which establishes an internal process for memory generation, retrieval, and prediction refinement through three components: Structured Camouflage State Memory (SCSM), Camouflage State Replay Adapter (CSRA), and Conflict-Guided Anchor Correction (CGAC).

Critical strengths

The paper’s strengths lie in its clear motivation; the three modules that address evidence organization, selective reuse, and conflict correction, respectively, form a cohesive methodological framework. Experiments conducted across four camouflaged object detection datasets demonstrate state-of-the-art (or tied-for-best) performance in 15 out of 16 evaluation metrics. The authors also provide ablation studies on the primary modules and their core sub-components, further supported by image-pair analysis, qualitative visualizations, and cross-task experiments, which effectively substantiate the method's overall efficacy and potential scope of application.

Critical weaknesses

A primary limitation is that the innovation leans more toward a task-specific combination of existing techniques, leaving the independent value of the structured memory mechanism insufficiently validated. Specifically, the paper lacks direct comparisons, i.e., under equivalent parameter counts and computational budgets, against standard two-stage refinement, iterative inference, or conventional prototype-based attention schemes; consequently, it is difficult to disentangle the contributions of the proposed mechanism itself from those of the additional computation. While Tables 3–5 cover the main sub-components, they do not independently verify the necessity of each "state" type, nor do they granularly analyze specific design elements such as support-versus-competition relationships or spatial/channel gating. Furthermore, the full configuration does not consistently outperform simplified versions across all metrics. It is recommended that the authors conduct additional controlled experiments and expand the scope of the ablation studies to include more datasets. For performance gains of smaller magnitude, the paper should report results from multiple independent training runs, along with confidence intervals and appropriate paired significance tests. Finally, the evaluation could be strengthened by including analyses of boundary quality, confidence calibration, or error-region distribution, as well as metrics for computational cost, inference latency, and memory footprint, to provide a more comprehensive assessment of boundary refinement, reliability, and practical efficiency.

Questions for authors

The authors are requested to clarify the following points: (1) Given the same backbone network, training configuration, and inference budget, what is the extent of the advantage offered by CamMem compared to standard two-stage refinement or conventional prototype-based attention methods? (2) What is the rationale behind the design of the five state categories, their respective weighting formulas, and the priors governing support and competitive relationships? Furthermore, how do the results change when these states are removed one by one, or when the relationship branches are disabled individually? (3) Are the primary state-of-the-art (SOTA) comparisons and ablation study results stable across multiple random seeds, and are the observed marginal performance differences statistically significant? (4) How much additional inference cost is incurred by memory retrieval and target-centric local refinement, respectively? Additionally, in cases where the initial prediction is severely erroneous or a target is missed entirely, is the framework capable of correcting these errors, rather than exacerbating them?

5 comments

We sincerely thank the reviewer for the thoughtful and constructive feedback. We particularly appreciate your recognition of CamMem's cohesive methodological framework, strong empirical performance, and comprehensive experimental evaluation. We also acknowledge your concerns regarding the independent contribution of structured memory, the necessity of individual memory states and relation mechanisms, the robustness of experimental results, and computational efficiency. To address these concerns, we have conducted additional comparisons with conventional refinement and prototype attention methods, state-specific and relation-branch ablations, multi-seed evaluations, and detailed analyses of inference overhead and error recovery. We also clarify the accuracy–efficiency trade-offs and the limitations of interpreting marginal performance differences without formal significance testing. We provide our detailed responses and additional evidence below.


Q1: Given the same backbone network, training configuration, and inference budget, what is the extent of the advantage offered by CamMem compared to standard two-stage refinement or conventional prototype-based attention methods?

A1: We thank the reviewer for this valuable suggestion. To assess the benefits of structured memory replay over conventional refinement, we introduce two additional baselines: Direct Two-Stage Refinement (DTR) and Foreground/Background Prototype Attention (FG/BG-PA), using the same SAM2 backbone and training configuration. For a focused comparison, we evaluate CamMem (w/o CGAC) using the replay prediction directly, without anchor correction or target-centric local refinement. This setting allows us to examine the joint effects of structured memory construction and replay.

R2-Table 1(a). Computational comparison with conventional refinement methods.

MethodTrainable Params (M)GFLOPsLatency (ms)Peak Memory (MiB)
DTR17.447905.2758.701125.91
FG/BG-PA17.329891.5660.951123.16
CamMem (w/o CGAC)17.433870.3496.171126.38

R2-Table 1(b). Quantitative comparison on NC4K and CHAMELEON.

DatasetMethodM↓Mean F↑Sm↑Ex↑
NC4KDTR0.0280.8610.9070.943
FG/BG-PA0.0280.8690.9070.952
CamMem (w/o CGAC)0.0270.8850.9090.947
CHAMELEONDTR0.0190.8650.9130.960
FG/BG-PA0.0200.8670.9160.963
CamMem (w/o CGAC)0.0180.8860.9220.956

Performance and computational analysis. As shown in R2-Table 1(b), CamMem (w/o CGAC) achieves the highest mean F-measure on both datasets, outperforming DTR and FG/BG-PA by 2.4 and 1.6 percentage points on NC4K, and 2.1 and 1.9 percentage points on CHAMELEON, respectively. It also achieves the highest S-measure on both datasets, although FG/BG-PA obtains higher maximum E-measure. Regarding computational efficiency, R2-Table 1(a) shows that CamMem requires fewer profiler-estimated GFLOPs and maintains comparable trainable parameter counts and peak GPU memory usage, but incurs higher inference latency (96.17 ms versus 58.70 and 60.95 ms). Therefore, while the backbone and training configurations are controlled, the inference budgets are not strictly matched. These results support the benefits of structured memory construction and replay over conventional refinement and prototype attention, while revealing an accuracy–latency trade-off.

We will incorporate these additional comparisons and computational analyses into the revised manuscript.

Q2: What is the rationale behind the design of the five state categories, their respective weighting formulas, and the priors governing support and competitive relationships? Furthermore, how do the results change when these states are removed one by one, or when the relationship branches are disabled individually?

A2: We thank the reviewer for this important question. We clarify the design rationale and provide additional state-specific and relation-branch ablations.

(1) State definitions and weighting rationale. The five states represent complementary evidence derived from the initial prediction. FG and BG use confidence-weighted probabilities to emphasize confident foreground and background responses. BD combines local prediction differences with ambiguity to highlight uncertain structural transitions. REL downweights boundary responses to retain reliable foreground evidence, while CONF aggregates uncertain foreground and boundary-associated conflict cues. These states are overlapping soft assignments, not mutually exclusive regions or ground-truth error labels. Their weighting formulas, provided in the manuscript, are heuristic inductive biases rather than calibrated error probabilities.

(2) Supportive and competitive priors. Supportive relations (e.g., FG–REL and BD–CONF) encourage interaction between complementary evidence states, whereas competitive relations (e.g., FG–BG, REL–CONF, and BG–REL) model opposing evidence. The corresponding priors bias content-dependent inter-state similarities before softmax, without imposing hard constraints. Learnable coefficients control supportive attraction and competitive separation in replay-value updates.

(3) State-specific aggregation ablation. Directly removing a state would change CSRA's fixed memory interface and state contrasts. Instead, we replace one state's weighted feature aggregation with global average pooling (GAP), retaining its state embedding, the other four states, and CSRA. This tests state-specific aggregation without deleting the state slot.

R2-Table 2: State-specific aggregation ablation on NC4K and CHAMELEON.

VariantNC4K M↓Mean F↑Sm↑Ex↑CHAMELEON M↓Mean F↑Sm↑Ex↑
FG → GAP0.0280.8770.9050.9420.0180.9000.9270.979
BG → GAP0.0270.8890.9090.9510.0180.8940.9250.966
BD → GAP0.0280.8840.9080.9490.0180.8960.9250.966
REL → GAP0.0270.8840.9110.9520.0190.8880.9230.971
CONF → GAP0.0270.8870.9100.9530.0180.9000.9270.979
Full CamMem0.0270.8900.9120.9530.0180.9020.9270.971

Full CamMem achieves the highest mean F-measure on both datasets. Replacing FG aggregation reduces mean F-measure from 0.890 to 0.877 on NC4K; replacing REL aggregation reduces it from 0.902 to 0.888 on CHAMELEON. Some replacements improve individual metrics (e.g., CHAMELEON Ex), so the findings support the overall benefit of state-specific aggregation, not universal gains from every state.

(4) Relation calibration ablation. We separately remove the supportive or competitive replay-value update term while keeping the five states, evidence-directed replay, spatial–channel gating, and CGAC unchanged. The relation-prior matrices are not independently ablated.

R2-Table 3: Ablation of supportive and competitive relation updates on NC4K and CHAMELEON.

VariantNC4K M↓Mean F↑Sm↑Ex↑CHAMELEON M↓Mean F↑Sm↑Ex↑
w/o Supportive Branch0.0280.8810.9070.9480.0190.8800.9180.959
w/o Competitive Branch0.0270.8870.9090.9520.0180.8950.9260.970
Full CamMem0.0270.8900.9120.9530.0180.9020.9270.971

Removing either update term lowers mean F-measure on both datasets, with larger reductions for the supportive term in these runs. These ablations assess the replay-value updates, not individual prior connections or prior weights.

We will incorporate the additional ablations and clarify these design choices in the revised manuscript.

Q3: Are the primary state-of-the-art (SOTA) comparisons and ablation study results stable across multiple random seeds, and are the observed marginal performance differences statistically significant?

A3: Thank you for this important question. To examine the robustness of CamMem, we conducted three independent training runs using different random seeds (42, 123, and 3407) under identical experimental settings. We report the mean and standard deviation (Mean ± Std) across these runs.

(1) Stability of SOTA performance. Across the three random seeds, the full CamMem achieves weighted F-measure scores of 0.8663 ± 0.0020, 0.8414 ± 0.0014, 0.8720 ± 0.0027, and 0.8888 ± 0.0005 on CAMO, COD10K, NC4K, and CHAMELEON, respectively. The small standard deviations indicate that CamMem maintains relatively stable performance across different random seeds.

(2) Stability of the main component ablations. To further assess the reliability of the component-level improvements, we repeated all four main ablation configurations using the same three random seeds. The results on NC4K and CHAMELEON are summarized in R2-Table 4.

R2-Table 4: Multi-seed stability analysis of the main component ablations on NC4K and CHAMELEON (Mean ± Std, three seeds).

NC4K

CSRASCSMCGACM ↓Mean F ↑Sm ↑Ex ↑
———0.0375 ± 0.00120.8458 ± 0.00420.8879 ± 0.00260.9294 ± 0.0031
✓——0.0286 ± 0.00160.8779 ± 0.00340.9064 ± 0.00180.9439 ± 0.0027
✓✓—0.0273 ± 0.00060.8831 ± 0.00290.9078 ± 0.00190.9461 ± 0.0020
✓✓✓0.0274 ± 0.00070.8876 ± 0.00270.9103 ± 0.00170.9520 ± 0.0014

CHAMELEON

CSRASCSMCGACM ↓Mean F ↑Sm ↑Ex ↑
———0.0307 ± 0.00140.8455 ± 0.00460.8923 ± 0.00280.9386 ± 0.0041
✓——0.0196 ± 0.00080.8653 ± 0.00380.9105 ± 0.00190.9501 ± 0.0034
✓✓—0.0186 ± 0.00050.8851 ± 0.00230.9218 ± 0.00130.9568 ± 0.0031
✓✓✓0.0171 ± 0.00050.8989 ± 0.00260.9272 ± 0.00020.9710 ± 0.0049

The results demonstrate consistent improvement trends in the mean performance across the component configurations. On NC4K, mean F-measure increases from 0.8458 for the baseline to 0.8779 with CSRA, 0.8831 with SCSM, and 0.8876 for the full model. Similar improvements are observed on CHAMELEON, where the corresponding scores increase from 0.8455 to 0.8653, 0.8851, and 0.8989. The relatively small standard deviations further indicate limited run-to-run variability, while the progressive improvements in mean performance support the effectiveness of the proposed components.

(3) Robustness of performance improvements. The multi-seed experiments are intended to assess the reproducibility of the observed improvements. As shown in R2-Table 4, the mean F-measure increases progressively across the component configurations on both datasets, with relatively small variations across runs. These results suggest that the improvements are not solely attributable to a particular random seed. Nevertheless, formal statistical significance cannot be established from the three-run summaries alone, particularly for marginal differences and comparisons with competing SOTA methods.

Reviewer_qzyVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CamMem adapts SAM2 to single-image camouflaged object detection by turning a first-pass prediction into five prototype “memory” states (foreground, background, boundary, reliable, conflict), replaying them into the encoder through a gated adapter, and then blending the initial prediction back in with a bounded, conflict-guided correction, reporting state-of-the-art results on four COD benchmarks and encouraging results on salient object, defocus blur, and forgery detection.

Critical strengths
  • Clear motivation and structure. The what-to-memorize, how-to-replay, where-to-correct decomposition is easy to follow, and the idea of building memory from the image’s own prediction fits single-image COD.
  • Strong headline numbers. CamMem is compared against 31 methods, including many recent SAM/SAM2-based ones. It shows clear margins over SAM2-UNet on CAMO and NC4K.
  • Thorough supplementary analysis. Beyond tables, the appendix includes PR/F-threshold curves, per-image distributions, and paired per-image comparisons, which show gains are not driven by a few images.
  • Conservative correction design. Bounded residual mixing in CGAC is a sensible way to avoid destabilizing the main prediction. Breadth. The same architecture is applied to three other dense-prediction tasks, and code is provided.
Critical weaknesses
  • The paper separates itself from other memory-based methods on one axis, where the memory comes from. It does not discuss how it differs in mechanism, and it does not compare against them experimentally. The introduction explains what kind of memory CamMem builds but never names a failure of feed-forward COD that memory specifically solves.
  • The paper does not say what the memory framing adds over prediction-guided refinement (SINet, PFNet, UGTR, SegMaR, CamoFormer), so it is unclear whether this is a new memory or a refinement method with memory vocabulary.
  • CamMem is distinguished from other memory-based methods only by where memory comes from, with no comparison of how memory is written, read, or used (e.g., against CamSAM2 and RefOnce prototypes).
  • The paper claims to be inspired by SAM2’s memory but never uses it or tests the obvious alternative of feeding the first-pass mask through SAM2’s memory encoder or mask-prompt path.
  • Table 1 omits RetroMem, RefOnce, RA-COD, CamSAM2, and SRR, and RetroMem in particular is a directly relevant image-level COD method. So this paper may need to compare with these methods to support the strength of the proposed method.
Questions for authors

Please see the weaknesses.

5 comments

Response to Reviewer qzyV

Comment:

We sincerely thank the reviewer for the thoughtful and constructive feedback. We particularly appreciate your recognition of CamMem's clear “what to memorize, how to replay, and where to correct” design, strong empirical performance, and comprehensive supplementary analyses. We also acknowledge your concerns regarding the motivation and mechanistic distinctions of CamMem from existing prediction-guided refinement and memory-based methods. To address these concerns, we have further clarified the underlying motivation and memory mechanisms, and conducted additional experiments comparing CamMem with conventional refinement strategies, SAM2's native memory and mask-prompt mechanisms, and recent memory-based COD methods. We provide detailed responses to each concern below and will incorporate these clarifications and additional experimental results into the revised manuscript.


Q1: The paper separates itself from other memory-based methods on one axis, where the memory comes from. It does not discuss how it differs in mechanism, and it does not compare against them experimentally. The introduction explains what kind of memory CamMem builds but never names a failure of feed-forward COD that memory specifically solves.

A1: We thank the reviewer for this insightful comment. We agree that the motivation and mechanistic distinctions of CamMem should be more clearly articulated.

(1) Motivation. A key limitation of feed-forward COD is that ambiguous foreground–background evidence may remain unresolved during feature refinement. In highly camouflaged scenes, weak foreground responses can lead to missed object regions, while background textures resembling the target may produce false positives. Although existing methods progressively refine features and predictions, they do not necessarily maintain an explicit representation of these competing cues for subsequent reassessment. CamMem addresses this limitation by organizing initial visual evidence into structured memory states and revisiting ambiguous representations through memory-conditioned feature replay.

(2) Mechanistic distinction. Existing memory-based camouflage perception methods can be broadly categorized into two types: temporal memory and historical memory. Temporal-memory methods exploit information across video frames. Specifically, CamSAM2 constructs object prototypes from high-quality historical observations to guide video camouflage segmentation, while SRR selects informative reference frames and integrates long-term reference information with short-term motion cues. Historical-memory methods, such as RetroMem, retrieve relevant knowledge accumulated across previous samples to support current-image inference. Although these methods differ in memory construction and retrieval, both primarily reuse visual evidence acquired beyond the current image. In contrast, CamMem introduces an intra-image memory writing–relation-calibrated replay–anchor correction mechanism. SCSM organizes initial visual evidence into five complementary camouflage-state prototypes. CSRA then models supportive and competitive relationships among these states and performs evidence-directed replay to recalibrate encoder features. Finally, CGAC selectively corrects the replay-conditioned prediction through bounded anchor correction. Thus, CamMem uses memory to preserve and reassess competing camouflage evidence rather than merely retrieve historical information.

(3) Experimental evidence. The component ablations in Tables 2–4 support the effectiveness of the proposed memory mechanism. On NC4K, introducing CSRA with unified memory improves mean F-measure from 0.848 to 0.881. On CHAMELEON, replacing unified memory with five structured states further improves mean F-measure from 0.868 to 0.886. Additional ablations examine the contributions of memory-state composition, relation calibration, evidence-directed replay, and spatial–channel gating. Furthermore, as detailed in our responses to Q4 and Q5, we provide additional comparisons with SAM2's native memory and mask-prompt mechanisms, as well as relevant memory-based COD methods. These experiments complement the mechanistic analysis, while comparisons involving different architectures and evaluation protocols are interpreted with appropriate caution.

Q2: The paper does not say what the memory framing adds over prediction-guided refinement (SINet, PFNet, UGTR, SegMaR, CamoFormer), so it is unclear whether this is a new memory or a refinement method with memory vocabulary.

A2: We thank the reviewer for this important comment. CamMem shares the coarse-to-fine objective of existing methods, but differs in how initial evidence is represented and reused.

(1) Mechanistic distinction. SINet separates camouflage perception into search and identification stages; PFNet progressively refines localization through distraction mining; UGTR uses uncertainty-guided transformer reasoning; SegMaR iteratively segments and magnifies candidate regions; and CamoFormer uses masked separable attention for progressive refinement. These methods establish the value of prediction guidance and repeated refinement, which we do not claim as novel.

In contrast, CamMem introduces an explicit memory writing–retrieval–replay mechanism within a single image. SCSM aggregates first-pass visual features using five overlapping, prediction-derived soft weights into foreground, background, boundary, reliable, and conflict prototypes. Rather than retaining only a coarse spatial prediction or attention mask, it preserves separately addressable feature-level representations with distinct evidence roles. These image-specific prototypes are reused during the second encoding pass.

CSRA retrieves the prototypes through feature–memory matching, calibrates supportive and competitive inter-state relationships, and constructs signed foreground–background and reliable–conflict corrections. Spatial–channel gates regulate their injection into encoder representations. Thus, CamMem remains a refinement framework at the task level, but implements refinement through structured evidence storage and relation-calibrated replay rather than prediction guidance alone. The comparisons below assess the empirical value of this design.

(2) Experimental evidence. We compare CamMem (w/o CGAC) with Direct Two-Stage Refinement (DTR) and Foreground/Background Prototype Attention (FG/BG-PA) using the same SAM2 backbone and training configuration. CamMem (w/o CGAC) directly outputs the replay prediction without anchor correction or target-centric local refinement, allowing us to examine the joint contributions of structured memory construction and replay.

R1-Table 1(a): Computational comparison with conventional refinement methods.

MethodTrainable Params (M)GFLOPsLatency (ms)Peak Memory (MiB)
DTR17.447905.2758.701125.91
FG/BG-PA17.329891.5660.951123.16
CamMem (w/o CGAC)17.433870.3496.171126.38

R1-Table 1(b): Quantitative comparison on NC4K and CHAMELEON.

DatasetMethodM↓Mean F↑Sm↑Ex↑
NC4KDTR0.0280.8610.9070.943
FG/BG-PA0.0280.8690.9070.952
CamMem (w/o CGAC)0.0270.8850.9090.947
CHAMELEONDTR0.0190.8650.9130.960
FG/BG-PA0.0200.8670.9160.963
CamMem (w/o CGAC)0.0180.8860.9220.956

Performance and computational analysis. As shown in R1-Table 1(b), CamMem (w/o CGAC) achieves the highest mean F-measure on both datasets, outperforming DTR and FG/BG-PA by 2.4 and 1.6 percentage points on NC4K, and 2.1 and 1.9 percentage points on CHAMELEON, respectively. It also achieves the highest S-measure on both datasets, although FG/BG-PA obtains higher maximum E-measure. As shown in R1-Table 1(a), CamMem has fewer profiler-counted GFLOPs and comparable trainable parameters and peak GPU memory, but incurs higher latency (96.17 ms versus 58.70 and 60.95 ms). Thus, although the backbone and training configurations are controlled, inference budgets are not strictly matched. The results indicate benefits beyond conventional refinement and prototype attention, with an accuracy–latency trade-off.

Q3: CamMem is distinguished from other memory-based methods only by where memory comes from, with no comparison of how memory is written, read, or used (e.g., against CamSAM2 and RefOnce prototypes).

A3: We thank the reviewer for highlighting this distinction. CamMem differs from CamSAM2 and RefOnce in what the prototypes represent, how they are constructed and queried, and how retrieved evidence modifies visual features.

(1) Memory writing. CamSAM2 constructs object prototypes by applying farthest-point sampling and clustering to features within predicted object regions, then stores them for subsequent video frames. RefOnce pools reference-object features and accumulates category-level prototypes through exponential moving average updates during training. CamMem instead uses state-weighted feature aggregation to construct five image-specific prototypes from initial perception. Their identities correspond to foreground, background, boundary, reliable, and conflict evidence, rather than object clusters or semantic categories. The overlapping assignments preserve distinct evidence roles, including background and ambiguous responses, for subsequent comparison.

(2) Memory reading. CamSAM2 uses current-frame features as queries and historical object prototypes as keys and values in its explicit object-aware fusion module. RefOnce predicts query-conditioned mixture weights over stored category prototypes to synthesize a guidance vector, without requiring reference images at inference. CamMem retains the five states as separately addressable memory entries. CSRA computes content-dependent inter-state similarities and combines them with supportive and competitive priors to calibrate replay values. Encoder features then query these calibrated memory states through feature–memory matching. This introduces explicit relationships among evidence states beyond query–prototype relevance matching.

(3) Memory utilization. CamSAM2 incorporates retrieved object details into feature fusion for video segmentation, whereas RefOnce uses bidirectional attention alignment to adapt query features and synthesized category guidance. In CamMem, CSRA reuses the state memory across second-pass SAM2 encoder blocks. Alongside relation-calibrated retrieval, signed FG–BG and REL–CONF contrasts generate directional feature corrections, whose injection is regulated by spatial–channel gates. CGAC subsequently performs bounded prediction-level anchor correction, which is separate from the memory-reading operation.

Overall, the distinction extends beyond memory provenance: CamSAM2 preserves object details across frames, RefOnce consolidates category guidance across references, and CamMem retains and compares complementary evidence states within a single image. We will make these mechanistic distinctions more explicit in the revised manuscript.

Comments on the paper

1 comment

Dear Reviewers and Area Chairs, 

We sincerely thank all reviewers for their constructive feedback and valuable suggestions. 

We have provided detailed point-by-point responses to all reviewer comments, along with additional experimental evidence, including controlled comparisons, component-level ablations, multi-seed evaluations, computational analyses, and clarifications of the training and evaluation protocols. 

We would greatly appreciate your consideration of these additional results and clarifications during the discussion process. We hope that our responses help address the key concerns raised in the reviews, and we would be happy to provide any further clarification if needed. 

Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.