{"id":"2afe0bc1-6e96-4a4d-89b1-eb494dd0ddd6","arxiv_id":"2608.06150","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"CogVis achieves state-of-the-art open-vocabulary change detection on seven remote-sensing benchmarks by computing a reusable category-agnostic change prior once per image pair and calibrating per-query decision thresholds from a memory bank.","lead":"CogVis splits open-vocabulary change detection into a shared scene-change perception step, a memory-based semantic calibration step, and a region verification step, so temporal evidence is computed once per image pair. It reports state-of-the-art accuracy on seven remote-sensing change benchmarks and a 28.5% throughput gain over the fastest prior method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'state-of-the-art across all seven datasets' claim is not yet established because the comparison mixes cited baseline numbers and single-run metrics; margins of +0.05 (LEVIR-CD) and +0.14 (WHU-CD) points are within plausible run-to-run noise.","rationale":"The reader's conditional verdict is well-founded. The strongest claim is empirical, and the empirical support has three weak spots: heterogeneous baseline sourcing, single-run reporting, and per-class blindness masked by macro mIoU. The ARF recall ceiling identified by the reader is real but secondary: it matters only if the candidate mask already drops true changes, and the xBD minor-damage F1 of 0.86% suggests this can happen. Re-running baselines under one protocol and reporting variance would settle the SOTA claim; if the margins survive, the conceptual contribution stands. The verdict therefore remains conditional rather than accept or reject.","tokens_in":22049,"tokens_out":5743,"duration_ms":62453,"concrete_test":"Create one evaluation harness for all methods: fixed 512x512 inputs, complete test splits, the exact prompt vocabulary of Tables S1-S3, and the class-wise metrics of Eq. S2. Run CogVis and the released baselines (UCD-SCM, AnyChange, DynamicEarth, AdaptOVCD, OpenDPR, OmniOVCD, Seg2Change, CoRegOVCD) for at least three seeds; report mean and standard deviation on LEVIR-CD, WHU-CD, and xBD, plus per-state F1 for xBD. If the LEVIR-CD/WHU-CD margins fall within one standard deviation, or if xBD minor-damage F1 remains below the best baseline, the abstract's 'state-of-the-art across all evaluated datasets' should be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim lands only if CogVis's margins are real and reproducible under a unified protocol. Table 1 reports gains of +0.05 on LEVIR-CD and +0.14 on WHU-CD, yet no error bars, seeds, or re-executed baselines are provided; the table appears to take competitor numbers from their original papers. The sensitivity analysis in Table 5 is performed on the same test benchmarks, and the default configuration is the best row, meaning the headline result may reflect test-set selection rather than a fixed inference rule. The xBD claim is likewise fragile: Table S7 shows minor-damage F1 of 0.86%, below OpenDPR's 3.26%, so the mIoU gain is driven by destroyed (50.83) and major (11.69) states; a damage-assessment 'state of the art' that essentially cannot detect minor damage is an overstatement. These are checkable empirical issues, not conceptual flaws, but they are the load-bearing supports for the abstract's SOTA statement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CogVis, an open-vocabulary change detection (OVCD) framework that reformulates the task as a perception–memory–verification pipeline. A Scene Change Perceptron (SCP) computes a query-shared, category-agnostic change prior from frozen bi-temporal features; a Semantic Memory Calibrator (SMC) retrieves memory entries to estimate an image-query-specific decision threshold; and an Adaptive Region Filter (ARF) verifies connected components using semantic, temporal, and geometric evidence. The adapters are trained sequentially on a held-out source dataset (CA-CDD), and the visual encoder, decoder, and CLIP text encoder remain frozen. Experiments on seven benchmarks report state-of-the-art mIoU/IoU on all of them, with gains from +0.05 to +8.17 points over prior methods, plus a 28.50% throughput improvement from sharing temporal perception across queries. The supplementary material provides class-resolved results, hyperparameter sensitivity, and a large-scene wildfire mapping case study.","tokens_in":22354,"tokens_out":2794,"duration_ms":30510,"significance":"If the empirical claims hold, CogVis provides a conceptually clean answer to a question implied by the title: OVCD need not recompute category-agnostic temporal evidence for every query. The design separates decision scopes and shows that a single change prior can support multiple text queries, which is a plausible and potentially useful architectural insight for remote sensing monitoring systems. The paper ships code and uses a disjoint source dataset for adapter training and memory construction, which is a strength for transferability. It also includes a thorough ablation study showing that each component contributes and that memory retrieval transfers an operating point rather than feature prototypes. The main significance rests on the claim of simultaneous accuracy and efficiency gains across all seven benchmarks; that claim is currently supported by single-run metrics against cited baselines, and the margins on two datasets are small enough that the result needs stronger empirical verification before it can be taken as established.","major_comments":[{"comment":"The abstract's claim that CogVis achieves state-of-the-art performance across all evaluated datasets is not yet established by the evidence in Table 1. The gains on LEVIR-CD (+0.05) and WHU-CD (+0.14) are well within typical run-to-run variation for trained change detectors, yet no error bars, number of seeds, or significance tests are reported. The baseline numbers are cited from their original papers rather than re-executed under a unified protocol with identical frozen features and evaluation code. To support the SOTA claim, the authors should provide variance estimates (e.g., at least three seeds) for their own method and, ideally, re-run the strongest baselines under the same inference and prompt vocabulary, or clearly state which numbers were taken from prior papers and justify why the comparison is fair without re-execution.","section":"Table 1 and Abstract"},{"comment":"The default configuration in Table 5 is the best-performing row on the same test benchmarks used for the headline results. For example, varying Kmem, Tm, tau_r, and eta_a all produce lower SECOND mIoU than the default, and the default also achieves the best CLCD IoU. If these hyperparameters were chosen by evaluating on the test set, the reported margins may reflect selection bias rather than a fixed inference rule. The paper should state clearly whether hyperparameters were selected on a validation split or by any test-set peeking, and if the latter, it should be acknowledged as a limitation. The robustness claim in the text ('CogVis is robust to memory-retrieval settings') is weakened because the ablation is performed on the same benchmarks that define the headline numbers.","section":"Table 5 and Section 'Sensitivity Analysis'"},{"comment":"The claim of state-of-the-art performance on building-damage assessment (xBD) is overstated when class-resolved results are considered. Table S7 shows that CogVis's minor-damage F1 is only 0.86%, below OpenDPR's 3.26%, and its no-damage F1 is only slightly better than AdaptOVCD (88.06 vs 87.46). The large mIoU gain is driven almost entirely by the destroyed class (50.83 vs 35.42 for DynamicEarth-MCI) and, to a lesser extent, major damage (11.69 vs 3.67). A damage-assessment model that essentially cannot detect minor damage, and is only marginally better than the strongest baseline on no damage, should not be summarized as achieving SOTA across the board without a clear qualification that the advantage is concentrated in severe damage states. The main paper's sentence 'the largest gain of 8.17 points' on xBD should be connected to this class-level pattern.","section":"Table S7 and xBD discussion"},{"comment":"The paper explicitly states 'ARF removes unreliable proposals but cannot recover regions missing from fMc.' This structural fact means the final prediction is upper-bounded by the recall of the SCP plus SMC front end. The paper does not report the recall of the candidate mask fMc before ARF, nor does it analyze how often true changes are suppressed by the change-prior support term G_c in Eq. (5) and Eq. (S12) or by the calibrated threshold tau_c in Eq. (7). Since the central claim is that the framework improves accuracy across all benchmarks, and since ARF can only filter, the reader needs to know whether the reported gains come from improving recall (via SCP/SMC) or from precision improvement via ARF. The ablation in Table 2 shows ARF improves results when combined with SCP and SMC, but does not decompose false negatives introduced upstream. Reporting front-end recall and the proportion of final true positives that originate from components that would have been rejected by a fixed threshold would directly address this load-bearing limitation.","section":"Methodology, 'Adaptive Region Filter'"}],"minor_comments":[{"comment":"The notation in Eq. (5) uses a product symbol 'Π' for the composition of the four cues, but it is not clear whether the product is elementwise over pixels and how the clipping to [0,1] interacts with the terms S_c^p_c. Please define the operation explicitly and ensure the dimensions of D_c, G_c, S_c, and p_c are consistent.","section":"Equation (5)"},{"comment":"Figure 2 is dense and the three adapter checkpoints are not clearly separated. Consider marking the frozen versus trainable components with a single consistent legend and labeling the sequential training order directly on the diagram.","section":"Figure 2"},{"comment":"The reference list contains duplicate entries for 'Chen, H.; and Shi, Z.' (2020a and 2020b), and the in-text citation '(Chen and Shi 2020b)' on LEVIR-CD is actually the journal paper while the dataset is cited as 2020a. Please reconcile the citation keys to avoid ambiguity.","section":"References"},{"comment":"The efficiency measurements are reported after model initialization with a batch size of one, but the query-scaling measurement in Section 'Efficiency and Scalability' mentions one warm-up and one synchronized full-request measurement. Please specify whether the reported FPS is the mean over multiple runs or a single run, and report the standard deviation if available.","section":"Section A.2, Eq. (S3)"}],"recommendation":"major_revision","confidential_remarks":"The paper is addressing a relevant problem and the conceptual framework is appealing, but the empirical evidence for the headline claim needs strengthening. The authors should be asked to provide error bars or multiple seeds, re-run or clearly justify the baseline numbers, and address the xBD overstatement. The structural upper-bound issue with ARF is worth a thoughtful response; it may be sufficient for the authors to report candidate recall and analyze failure modes. I do not see a fundamental correctness flaw, so rejection is not appropriate at this stage. The scope fits an AI/remote-sensing venue well, and the code release is a positive signal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The gist: CogVis is a genuine reformulation of OVCD, not a bag of tricks. Splitting the pipeline into shared change perception, memory-transferred thresholds, and region verification makes sense, and the memory bank of operating points rather than feature prototypes is a real departure from MemOVCD and friends. The paper is also unusually honest: it states plainly that ARF cannot recover candidates missing from the front end, which is the structural limit you'd want to know. Ablations and the CA-CDD source-dependence test are careful, and the efficiency analysis is transparent.\n\nThe soft spots are all about the headline numbers. The 'state-of-the-art across all seven datasets' claim is load-bearing and it is not yet established. There are no error bars; margins of +0.05 on LEVIR-CD and +0.14 on WHU-CD are inside run-to-run noise. Baselines are cited from their papers rather than re-run under one protocol, so some gaps may be protocol artifacts. On xBD the mIoU gain is real but concentrated in destroyed and major classes; minor-damage F1 is 0.86%, below OpenDPR's 3.26%, so 'state of the art' for building-damage assessment is an overstatement unless you restrict the claim to the macro metric. Sensitivity analysis is run on the same test benchmarks with the default configuration as the best row, so the default may have been chosen on the test set. These are checkable empirical issues, not conceptual ones.\n\nNet: the architecture and the factorization are solid and worth engaging with. The paper deserves a serious referee. What it needs is a revised reporting protocol: multiple seeds, re-executed baselines, per-class xBD numbers in the main text, and a sensitivity analysis on a validation split or with the default fixed before test evaluation.","headline":"CogVis is a real architectural contribution to OVCD, but the 'SOTA on all seven benchmarks' claim is not yet backed by the numbers as reported.","tokens_in":22862,"tokens_out":1927,"would_cite":true,"duration_ms":19249,"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":"Open-vocabulary change detection need not perceive the scene anew per query: shared temporal evidence, a memory-calibrated threshold, and a region filter beat rivals on seven benchmarks with 28.50% higher throughput.","keywords":["open-vocabulary change detection","remote sensing","perception-memory-verification","semantic memory calibration","scene change perceptron","adaptive region filter","multi-query efficiency","foundation models"],"falsifier":"Measure the recall of the candidate mask $\\hat{M}_c$ (the response thresholded at $\\tau_c$, before the Adaptive Region Filter) against ground truth on the seven benchmarks; if any query-relevant change class shows substantial false negatives at this stage, the published final IoU is capped by that recall. A direct version: construct a bi-temporal pair whose ground-truth change produces no detectable difference in the frozen backbone features, for instance a rebuilt identical structure or gradual vegetation stress, and check whether any threshold or filter recovers it — the paper's premise implies it cannot.","tokens_in":21816,"feed_emoji":"🛰️","tokens_out":15526,"duration_ms":125843,"temperature":0.7,"pith_summary":"The paper asks whether open-vocabulary change detection must re-analyze the whole scene for every text query, and answers no. It claims that the three decisions hidden in this task — what changed temporally, whether the change matches the queried category, and whether a candidate region is reliable — operate at different scopes and should not be entangled in a fixed cascade. CogVis computes a category-agnostic change prior once per image pair, calibrates a query-specific decision threshold by retrieving past image–query cases from a memory of operating points, and filters candidate regions by learned semantic, temporal, and structural reliability. On seven benchmarks spanning semantic change detection, binary change localization, and building-damage assessment, the paper reports state-of-the-art results on every dataset and a 28.50% inference-throughput gain from sharing the temporal perception step across queries.","feed_headline":"Shared scene scan cuts compute 28.5% and tops all 7 change benchmarks","feed_subtitle":"Decoupling temporal evidence from per-query semantics raises accuracy while cutting repeated computation by 28.5%","key_machinery":"The carrying object is the decoupling itself, encoded in the product rule $R_c = \\Pi_{[0,1]}(D_c \\odot G_c \\odot S_c \\cdot p_c)$: the per-query semantic response is the agreement of four cues, and one of them, $G_c = \\epsilon_g + (1-\\epsilon_g)B$, injects the shared change prior $B$ so that query-specific semantics inherit temporal evidence without re-analyzing the image pair. The second load-bearing mechanism is threshold transfer: the memory bank stores, for each source image–query pair, a key vector and an oracle threshold (the value that maximizes IoU over a grid), and retrieval combines the top-16 entries by temperature-weighted cosine similarity into an anchor $\\bar{\\tau}_c$, which a small Score Adapter corrects to $\\tau_c$. The third mechanism is region-level verification: a 19-dimensional descriptor per connected component (geometry, semantic confidence, temporal support, calibrated margin) feeds a Gate Adapter whose reliability score $u_j$ decides, together with a large-area protection rule, whether the component survives.","core_discovery":"The central claim is that the instability and redundant computation of open-vocabulary change detection come from coupling decisions with different scopes, and that decoupling them improves accuracy and efficiency at once. Temporal evidence is shared across all queries, so CogVis extracts it once with a Scene Change Perceptron that turns frozen bi-temporal features (concatenation, absolute difference, elementwise product, plus a gated low-rank residual) into a change prior $B$. For each query $c$, a Semantic Memory Calibrator builds a semantic-change response $R_c = \\Pi_{[0,1]}(D_c \\odot G_c \\odot S_c \\cdot p_c)$ — the agreement of semantic transition, support from the shared prior ($G_c = \\epsilon_g + (1-\\epsilon_g)B$), suppression of stable regions, and change in query presence — and sets a threshold $\\tau_c = \\Pi_{[\\tau_{\\min},\\tau_{\\max}]}(\\max\\{\\bar{\\tau}_c + \\Delta\\tau_c,\\ 0.75\\,T_{\\mathrm{adp}}(R_c)\\})$ from a temperature-weighted memory anchor and a learned correction. An Adaptive Region Filter then keeps each connected component only if it is large or passes a learned reliability gate. The paper's experiments claim this perception–memory–verification factorization is state of the art on all seven evaluated datasets, with the largest gain (8.17 mIoU points) on the building-damage benchmark.","pith_inferences":["My inference: the same perception–memory–verification split could transfer to other open-vocabulary spatio-temporal tasks, such as open-set video anomaly detection or multi-label land-use monitoring, wherever one shared 'what changed' signal must answer many textual queries.","My inference: because the memory stores thresholds rather than visual prototypes, a deployment could refresh the bank with operating points from new sites or sensors without retraining the frozen backbone, turning domain adaptation into a lookup-table update.","My inference: the paper's premise predicts a specific failure mode — on slow-onset or spectrally subtle changes that the frozen RGB encoder registers as no temporal difference, the candidate mask is blind and no later stage can help; the strongest published gains on sudden, high-contrast damage are consistent with this boundary.","My inference: a direct test of the reuse claim is the slope of latency versus vocabulary size; if the shared prior dominates, per-query marginal cost should stay roughly constant as $K$ grows, while per-query re-analysis baselines should scale with full scene complexity."],"forward_implications":["Multi-query cost becomes additive: latency follows $T(K) = T_{\\mathrm{pair}} + K\\,T_{\\mathrm{query}}$, so a vocabulary of ten queries costs seconds rather than ten full scene parses.","The change prior can be improved independently of semantic accuracy: any better category-agnostic change detector could replace the Scene Change Perceptron without touching semantic decoding or the memory bank.","The memory of operating points is compact and transferable: a 10% memory bank retains roughly 95% of full-memory gains, so calibration is a small set of representative response–threshold relationships rather than stored imagery.","Nuisance robustness is where the decoupling pays most: the largest single gain (8.17 mIoU on xBD) and the strong improvements on CLCD and DSIFN are consistent with temporal evidence blocking shadows, seasonal shifts, and misregistration before semantics are consulted.","The Adaptive Region Filter is a verification stage, not a post-processor: ablations show it helps only when the change prior and the calibrator are both present, so the gains depend on the cascade order, not on any single module."],"supporting_citations":[{"why":"Supplies the frozen SAM3 visual encoder and prompt decoder whose feature pyramid SCP reuses and whose semantics SMC decodes.","marker":"Carion et al. 2025"},{"why":"Provides the CLIP embeddings used for text-prompt encoding and for the retrieval keys that pull memory thresholds.","marker":"Radford et al. 2021"},{"why":"Formalizes the M-C-I and I-M-C cascades that CogVis replaces and supplies the comparison baselines for the open-vocabulary paradigm.","marker":"Li et al. 2026b"},{"why":"Contributes the Seg2Change baseline and the CA-CDD source domain used to train the three adapters and build the memory bank.","marker":"Su et al. 2026"},{"why":"Provides the SECOND benchmark and its class-involvement protocol used as the primary semantic change detection test.","marker":"Yang et al. 2022"},{"why":"Provides the xBD building-damage assessment benchmark where CogVis reports its largest gain.","marker":"Gupta et al. 2019"},{"why":"Provides the SCSCD multi-class semantic change detection benchmark with its seven foreground prompts.","marker":"Tan et al. 2025"}],"fun_headline_variants":["One shared scan serves every query, cutting compute 28.5%","Decoupling temporal and semantic boosts all 7 change benchmarks","No need to perceive anew: shared prior cuts compute 28.5% and wins all 7","CogVis reuses one scene prior, wins 7 change tests, 28.5% faster","Shared scene prior answers every query: 28.5% faster, tops 7 benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated by the paper as 'ARF removes unreliable proposals but cannot recover regions missing from fMc,' is that the early stages never lose a true change: if the shared change prior or the calibrated threshold suppresses a real change, no later stage can bring it back, so the final result can never exceed the recall of the candidate mask.","fun_headline_variants_meta":{"raw":{"variants":["One shared scan serves every query, cutting compute 28.5%","Decoupling temporal and semantic boosts all 7 change benchmarks","No need to perceive anew: shared prior cuts compute 28.5% and wins all 7","CogVis reuses one scene prior, wins 7 change tests, 28.5% faster","Shared scene prior answers every query: 28.5% faster, tops 7 benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001221,"raw_usage":{"total_tokens":5084,"prompt_tokens":1067,"completion_tokens":4017,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":683,"completion_tokens_details":{"reasoning_tokens":3907}},"tokens_in":683,"tokens_out":4017,"duration_ms":29686,"temperature":1.0,"reasoning_tokens":3907,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:55:59.774699+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the recall of the candidate mask $\\hat{M}_c$ (the response thresholded at $\\tau_c$, before the Adaptive Region Filter) against ground truth on the seven benchmarks; if any query-relevant change class shows substantial false negatives at this stage, the published final IoU is capped by that recall. A direct version: construct a bi-temporal pair whose ground-truth change produces no detectable difference in the frozen backbone features, for instance a rebuilt identical structure or gradual vegetation stress, and check whether any threshold or filter recovers it — the paper's premise implies it cannot.","supporting_citations":[{"cited_title":"IEEE Transactions on Geoscience and Remote Sensing , volume =","cited_arxiv_id":null,"evidence_quote":"Provides the SECOND benchmark and its class-involvement protocol used as the primary semantic change detection test."},{"cited_title":"ISPRS Journal of Photogrammetry and Remote Sensing , volume =","cited_arxiv_id":null,"evidence_quote":"Provides the SCSCD multi-class semantic change detection benchmark with its seven foreground prompts."}],"review_version":1}