{"id":"cc28086b-98ac-4357-b917-380563314e2e","arxiv_id":"2501.04934","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A plug-and-play module that separates merged changed instances in weakly-supervised change detection, improving accuracy across seven baselines and five datasets.","lead":"The paper introduces a plug-and-play module, DISep, that helps weakly-supervised change detection models in remote sensing separate densely packed changed objects instead of merging them into one blob. It works by finding likely changed regions, dividing them into separate instances, and training pixels of each instance to share similar features, improving accuracy on five benchmark datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The instance-separation mechanism relies entirely on a high-pass CAM threshold; Table 4 shows the method fails when both thresholds equal the CAM score, and the fixed 0.60/0.40 settings are only tested on LEVIR-CD.","rationale":"I read the paper in good faith: the method is clearly described, the ablations are informative, and the code is available. However, the central contribution is a mechanism for separating dense instances, and that mechanism has a single point of failure: the high-pass threshold Th. The separation loss cannot create separation by itself (the authors even note that inter-instance contrast hurts), so the entire benefit rests on the thresholding/connectivity step isolating instances. Table 4 is the most telling: with Th=Tl=0.45 (the same level used for pseudo-labels), F1 falls below baseline, meaning the extra loss is actively counterproductive when the masks do not separate instances. This is not a hypothetical failure mode; dense instances in remote sensing frequently touch or share boundaries, and the paper provides no evidence that a fixed 0.60 threshold reliably separates them across datasets. The reader flagged exactly this assumption; I agree. I also note the absence of any instance-level quantitative metric, which leaves the 'lumping' claim unverified at the object level. My proposed threshold sweep across all datasets would directly test whether the default settings generalize. I do not think this warrants rejection—the method may work well with appropriate thresholds—but it must be demonstrated. Hence I keep the CONDITIONAL verdict unchanged.","tokens_in":19640,"tokens_out":6580,"duration_ms":64904,"concrete_test":"Run the Table 4 threshold sweep (Th in {0.45,0.50,0.55,0.60,0.65} × Tl in {0.35,0.40,0.45}) for TransWCD on all five datasets, reporting F1/IoU for each combination. Determine whether the optimal (Th,Tl) differs across datasets; if the optimal Th varies by more than 0.05, or if no threshold beats the baseline on any dataset, the fixed 0.60/0.40 default is not a general plug-and-play setting. Also compute the fraction of ground-truth instance pairs that are merged in the Th=0.60 mask; a high merge rate would directly contradict the mechanism's ability to separate dense instances.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DISep's instance retrieval (Sec. 3.3) performs connectivity search on the binary mask Mc obtained by thresholding CAM at Th (Eq. 2). If two changed instances touch in this mask, they receive one instance ID; the separation loss (Eq. 8) then pulls all their pixels toward a common centroid, actively reinforcing the lumping the method claims to fix. The paper's fixed Th=0.60 and Tl=0.40 were chosen based on a single ablation on LEVIR-CD (Table 4), where setting both to the CAM threshold 0.45 drops F1 to 58.72, below the 60.08 baseline. This shows the mechanism is not robust to threshold misspecification and can harm performance. Because no instance-level metric (e.g., number of separated objects) is reported, the central claim of reducing instance lumping is only indirectly supported by pixel-level F1/IoU gains. The plug-and-play generality across datasets with differing instance densities is therefore not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DISep, a plug-and-play training-time module for weakly supervised change detection (WSCD). It localizes changed-instance candidate regions by thresholding class activation maps at a high-pass threshold, assigns instance IDs via 8-neighborhood connectivity search, and then applies a separation loss that clusters pixels of each instance toward that instance's feature centroid. The module is added to the training objective of existing WSCD methods without changing inference. The authors evaluate DISep on top of seven WSCD baselines (three Transformer-based and four ConvNet-based) across five datasets and report consistent F1/IoU/OA improvements, including headline gains of +6.27 F1 and +6.44 IoU for TransWCD on LEVIR-CD. They also show that DISep can improve two fully supervised change detection methods.","tokens_in":19783,"tokens_out":7479,"duration_ms":76032,"significance":"If the central claim holds, DISep is a genuinely useful, low-cost module for a real failure mode in WSCD: instance lumping in dense scenes. The strengths of the paper are the broad experimental coverage (seven baselines, five datasets), the consistent positive deltas in nearly all settings, the ablation study over loss weight, thresholds, loss functions, and sampling scopes, the negligible inference overhead, and the release of code. These are substantive. The concern is that the specific quantity the method claims to improve—instance-level separation and the ability to count changed objects—is never directly measured, and the core mechanism relies on a hand-set CAM threshold whose misspecification can make the method perform worse than the baseline, as shown in Table 4. With additional instance-level evaluation and a more careful treatment of threshold sensitivity, the paper would provide stronger support for its claims.","major_comments":[{"comment":"The motivation and title center on instance lumping and on accurately quantifying the number of changed objects, but all reported metrics are pixel-level F1, OA, and IoU. No instance-level metric is reported: for example, the precision/recall of detected changed instances, the number of connected components before and after DISep, or split/merge errors. As a result, the experiments demonstrate pixel-level accuracy gains but do not directly demonstrate that instances are actually separated or that counting improves. I recommend adding an instance-level evaluation on at least LEVIR-CD and WHU-CD (using connected components of the binary change ground truth as instance proxies), or softening the claims about instance separation and quantification.","section":"Section 4.2, Tables 1–2"},{"comment":"The instance retrieval step is load-bearing: it performs connectivity search on the high-pass thresholded mask Mc. If two changed objects touch after thresholding, they receive the same instance ID, and the separation loss in Eq. (8) then pulls their pixels toward one centroid, potentially reinforcing the lumping the method is designed to fix. Table 4 shows that this is not merely hypothetical: setting both Th and Tl to 0.45 drops F1 from 60.08 to 58.72, below the baseline, and setting Th=0.50/Tl=0.40 also underperforms the selected configuration. The chosen Th=0.60/Tl=0.40 is ablated only on LEVIR-CD and then applied uniformly to all five datasets. Given the different instance densities across datasets (Fig. 2), the authors should either justify the transferability of the fixed thresholds, provide per-dataset sensitivity analyses, or design an adaptive threshold scheme.","section":"Section 3.3, Eq. (2), Table 4"},{"comment":"The separation loss only enforces intra-instance compactness; it contains no term that pushes different instances apart. Any separation achieved therefore comes from the connectivity structure already present in the high-passed CAM. The paper's own ablation in Table 5 shows that adding a centroid-to-centroid contrast term reduces F1 by 2.10%, which raises the question of whether an explicit inter-instance separation term is viable. The claim in Section 3.4 that close instance centroids are 'not a concern' is unsupported by any quantitative analysis. I would like the authors to analyze whether the method can ever correct an instance merge that survives the high-pass threshold, and to provide evidence—for example, feature-distance statistics before and after DISep—that the loss genuinely separates merged components.","section":"Section 3.4, Eq. (8), Section 4.3.4"},{"comment":"The abstract and Section 4.2 claim 'state-of-the-art performance', but Tables 1–2 only compare each of the seven methods with and without +DISep. There is no comparison to external WSCD methods that are not used as backbones; for instance, CS-WSCDNet is discussed in Section 2.1 but does not appear in any experiment. Without such comparisons, the SOTA claim is not supported by the presented evidence. I recommend either adding external baseline results on at least one dataset per table, or revising the wording to 'consistently improves existing WSCD methods'.","section":"Abstract and Section 4.2"}],"minor_comments":[{"comment":"The text reports WCDNet's WHU-CD IoU improvement as +8.24%, but Table 1 shows +6.24% (26.34 vs. 20.10); the text and table should be aligned.","section":"Section 4.2.1"},{"comment":"The notation C = CAM_i / Max(CAM_i) is unclear because the subscript i appears only on the numerator and the sum in the definition of CAM is over the channel dimension; please define the normalization over the spatial extent of the activation map and use consistent indexing.","section":"Eq. (1)"},{"comment":"The layout of Table 4 is difficult to read because Th and Tl values are paired in a single column without grouping; consider using separate columns with clear row grouping or spacing to show the two-dimensional threshold grid.","section":"Table 4"},{"comment":"The reference list contains the same paper by Shi et al. twice, once as 'Shi et al., 2021' and once as 'Shi et al., 2022', with the same DOI; this duplicate should be consolidated.","section":"References"},{"comment":"The t-SNE visualizations support the claim of improved feature grouping but are qualitative; adding a simple quantitative cluster metric (e.g., average intra-instance vs. inter-instance feature distance, or silhouette score) would make the evidence more convincing.","section":"Section 5.1, Figures 8–9"}],"recommendation":"major_revision","confidential_remarks":"This is a well-executed empirical study with a useful negative result in the ablation: the method can fail when the threshold is misspecified. The main gap is that the paper's central claim is about instance separation, yet no instance-level evaluation is provided. The 'state-of-the-art' language should be toned down or supported by external baselines. In its current form, the paper is publishable after a major revision that addresses threshold robustness and narrows the claims to what the pixel-level metrics actually support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it's the first WSCD method I've seen that directly targets the instance-lumping problem in dense scenes, and it does so with a simple, training-only module that adds negligible cost. Second, the paper is honestly evaluated: seven baselines, five datasets, ablations over the loss designs, thresholds, and sampling scopes. That breadth is the paper's real strength.\n\nWhat's new: a high-pass CAM threshold extracts instance localization masks, connectivity search assigns instance IDs, and a pixel-to-centroid separation loss pulls same-instance pixels together in feature space. The components are individually known from weakly supervised instance segmentation and metric learning, but their combination for WSCD is new. The gains are consistent across the board, for instance +6.27 F1 and +6.44 IoU for TransWCD on LEVIR-CD. Code is released. The extension to fully supervised change detection also shows small positive gains.\n\nSoft spots, in proportion. The stress-test note is right: the instance retrieval relies entirely on the high-pass threshold. If two touched instances get merged in the binary mask, the separation loss actively reinforces that merge by pulling their features to a common centroid. Table 4 shows this: setting both thresholds to 0.45 (the CAM score) drops F1 to 58.72, below the 60.08 baseline. So the mechanism is not robust to threshold misspecification, and the fixed 0.60/0.40 comes from a single ablation on LEVIR-CD. The paper claims this pattern repeats on other datasets but doesn't show the threshold sensitivity for them. Also, no error bars or significance tests anywhere. Given the typical run-to-run variance in this kind of training, some of the smaller deltas are probably not reliable. And the SOTA comparison omits the SAM-based CS-WSCDNet, which the authors cite in related work—that's an unfair comparison if CS-WSCDNet is a plausible baseline. The circularity concern is real but not fatal: the separation loss uses instance masks from the same model's CAMs, but the headline claim is measured against external baselines, so the improvement isn't defined into existence.\n\nWho this is for: researchers working on weakly supervised change detection or on using CAM-based instance cues in dense remote sensing scenes. A generalist won't find the technical novelty deep, but the module is practical and well tested. It deserves a serious referee. My recommendation: send it to review, and ask the authors to add error bars, report instance-level separation metrics (e.g., number of separated objects or split/merge errors), and discuss or include the SAM-based baseline. The threshold fragility should be analyzed more thoroughly, but it's addressable.","headline":"A solid plug-and-play module for reducing instance lumping in weakly supervised change detection, with consistent gains across many baselines, but the core mechanism shows threshold fragility and the evaluation lacks error bars.","tokens_in":20367,"tokens_out":1516,"would_cite":true,"duration_ms":17415,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a training-time instance-separation module to weakly supervised change detection splits densely packed changed objects that are incorrectly merged, improving pixel-level accuracy across seven weak-supervision baselines and five…","keywords":["weakly supervised change detection","instance lumping","class activation maps","dense instances","separation loss","connectivity search","remote sensing","plug-and-play module"],"falsifier":"Take a synthetic pair of bi-temporal images with two changed rectangles that touch at a single corner, so that the CAM at threshold $T_h = 0.60$ cannot separate them into two connected components. Train a baseline plus DISep and compare the number of predicted change instances: if the module labels both rectangles as one instance and the loss then pulls their pixel features together, the predicted change map will still be merged, demonstrating that the claimed separation is limited to objects that the threshold already disjoins.","tokens_in":19364,"feed_emoji":"🛰️","tokens_out":4975,"duration_ms":46312,"temperature":0.7,"pith_summary":"The paper proposes DISep, a plug-and-play training module for weakly supervised change detection (WSCD), where only scene-level change labels are available. Its central claim is that dense clusters of changed objects are routinely merged in pixel-level predictions, and that a three-step module—localize instance candidates via a high-pass threshold on class activation maps, group pixels into instance IDs by connectivity search, and enforce intra-instance pixel-feature consistency with a separation loss—reliably unpicks this lumping. Across three Transformer-based and four ConvNet-based baselines on the LEVIR-CD, WHU-CD, DSIFN-CD, SYSU-CD, and CDD datasets, DISep consistently improves F1 and IoU, e.g., +6.27 F1 and +6.44 IoU for TransWCD on LEVIR-CD, while adding negligible training cost and no inference cost. If correct, the result matters because accurate instance separation is what makes change quantification—counting damaged buildings, new structures, or lost vegetation—possible under cheap scene-level annotation.","feed_headline":"Plug-in module unpacks lumped change detections, +6.3 F1","feed_subtitle":"Scene-level labels only: DISep separates dense change instances across five datasets and seven baselines, with no inference cost.","key_machinery":"The module is a three-step loop around any WSCD classifier. First, Instance Localization thresholds the class activation map $C$ at a high-pass value $T_h$ to produce a binary mask of reliably changed pixels; unchanged background below $T_l$ is treated as a single unchanged instance. Second, Instance Retrieval runs an 8-neighborhood connectivity search over that mask, assigning every connected component a distinct instance ID $k$ and producing an instance identity mask $M_{id}$. Third, Instance Separation adds a separation loss $L_{\\text{sep}} = l_{pc} + l_{puc} + l_{pu}$, where each term is the mean squared Euclidean distance from pixels in an instance to the instance's centroid feature computed on the last-layer feature map $F$, covering changed instances in changed images, unchanged background in changed images, and whole unchanged images. The loss is weighted by $\\alpha = 0.1$ and added to the classification cross-entropy; the whole loop is iterative, so the instance IDs refresh as the features improve.","core_discovery":"DISep's central discovery is that instance identity, normally unavailable under scene-to-pixel supervision, can be manufactured from the same classification signals already used for change localization, and that using those manufactured IDs to cluster pixel embeddings sharpens change boundaries. The paper states that it 'achieves state-of-the-art performance by enhancing three Transformer-based and four ConvNet-based methods' on all five datasets, with consistent gains in F1, IoU, and overall accuracy for every baseline tested. The mechanism is training-time-only: once the model is trained, DISep's losses are removed and inference is unchanged, so the improvement comes without any architectural modification or added runtime.","pith_inferences":["I would expect the fixed threshold $T_h = 0.60$ to become a robustness bottleneck: the paper's own ablation shows that threshold choice flips the method from its best F1 to below baseline, so an adaptive per-image or per-object threshold is a natural next test.","The idea of manufacturing instance IDs by connectivity search on weakly supervised CAMs and then clustering pixel embeddings is not change-detection-specific; the same loop could sharpen weakly supervised instance segmentation of buildings, roads, or other remote-sensing objects.","The failure of centroid-wise contrast suggests DISep's separating power comes from compactifying each instance's features rather than pushing instances apart, which implies the method should be stress-tested on scenes with extremely wide instance size disparity, where intra-instance compactness alone may be insufficient."],"forward_implications":["Existing WSCD pipelines can be upgraded by adding DISep's three loss terms during training, gaining roughly +2 to +8 F1 points across ConvNet and Transformer baselines without any inference-time change.","Accurate counting of changed objects becomes feasible under scene-level supervision, since separated instance predictions let users count damaged buildings or new structures instead of reporting merged blobs.","The same instance-separation mechanism transfers to fully supervised change detection, improving SNUNet by +1.26 F1 and CTD-Former by +1.32 F1 on LEVIR-CD.","Because the module acts purely on pixel-to-centroid clustering of the model's own features, it is architecture-agnostic and could be inserted into other dense prediction heads that suffer from object merging."],"supporting_citations":[{"why":"Baseline method TransWCD; the paper's headline gains (+6.27 F1, +6.44 IoU on LEVIR-CD) are measured on top of it.","marker":"Zhao et al., 2023"},{"why":"ConvNet baseline WCDNet; large gains (+7.91 F1 on WHU-CD) demonstrate the method works on traditional CNN architectures.","marker":"Andermatt and Timofte, 2020"},{"why":"Supplies Class Activation Maps, the signal from which DISep obtains instance localization.","marker":"Zhou et al., 2016"},{"why":"Shows peak instance responses in CAMs are usable for instance localization, the premise of the high-pass thresholding step.","marker":"Zhou et al., 2018"},{"why":"Introduces the WHU-CD dataset, one of the five evaluation benchmarks with dense building changes.","marker":"Ji et al., 2018"},{"why":"Introduces the LEVIR-CD dataset, the main benchmark where 79.53% of image pairs contain multiple change instances.","marker":"Chen and Shi, 2020"}],"fun_headline_variants":["Plug-in separates dense change instances, boosts F1 by 6.3","Manual instance IDs from class maps sharpen change detection","Training-only trick: separate instances to fix change detection","Weak labels, strong separation: DISep gains SOTA on 5 datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The success of DISep hinges on one premise: the high-pass threshold applied to the CAM actually splits densely packed changed objects into separate regions; if two objects touch in the thresholded mask they are merged into one instance ID, and the separation loss then actively pushes their pixels to look alike, deepening the very lumping the method is meant to cure.","fun_headline_variants_meta":{"raw":{"variants":["Plug-in separates dense change instances, boosts F1 by 6.3","Manual instance IDs from class maps sharpen change detection","Training-only trick: separate instances to fix change detection","Weak labels, strong separation: DISep gains SOTA on 5 datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1340,"prompt_tokens":1018,"completion_tokens":322,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":250}},"tokens_in":634,"tokens_out":322,"duration_ms":3705,"temperature":1.0,"reasoning_tokens":250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:20:52.031252+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic pair of bi-temporal images with two changed rectangles that touch at a single corner, so that the CAM at threshold $T_h = 0.60$ cannot separate them into two connected components. Train a baseline plus DISep and compare the number of predicted change instances: if the module labels both rectangles as one instance and the loss then pulls their pixel features together, the predicted change map will still be merged, demonstrating that the claimed separation is limited to objects that the threshold already disjoins.","supporting_citations":[{"cited_title":", author Timofte, R","cited_arxiv_id":null,"evidence_quote":"ConvNet baseline WCDNet; large gains (+7.91 F1 on WHU-CD) demonstrate the method works on traditional CNN architectures."},{"cited_title":", author Khosla, A","cited_arxiv_id":null,"evidence_quote":"Supplies Class Activation Maps, the signal from which DISep obtains instance localization."},{"cited_title":", author Zhu, Y","cited_arxiv_id":null,"evidence_quote":"Shows peak instance responses in CAMs are usable for instance localization, the premise of the high-pass thresholding step."},{"cited_title":", author Shi, Z","cited_arxiv_id":null,"evidence_quote":"Introduces the LEVIR-CD dataset, the main benchmark where 79.53% of image pairs contain multiple change instances."}],"review_version":1}