- 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
- 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.
- 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
- 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.
- 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
- 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?
- 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.
- 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?
- 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.
- 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
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.
| CSRA | SCSM | CGAC | DUTS-TE M↓ | Max F↑ | Sm↑ | Em↑ | PASCAL-S M↓ | Max F↑ | Sm↑ | Em↑ |
|---|---|---|---|---|---|---|---|---|---|---|
| — | — | — | 0.025 | 0.933 | 0.918 | 0.947 | 0.050 | 0.897 | 0.860 | 0.905 |
| ✓ | — | — | 0.022 | 0.936 | 0.927 | 0.955 | 0.047 | 0.907 | 0.883 | 0.920 |
| ✓ | ✓ | — | 0.019 | 0.944 | 0.935 | 0.960 | 0.044 | 0.912 | 0.894 | 0.935 |
| ✓ | ✓ | ✓ | 0.018 | 0.947 | 0.939 | 0.962 | 0.042 | 0.916 | 0.899 | 0.933 |
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.
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.
(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.