{"id":"e7e17b3a-d0a0-452c-981a-566058a45eef","arxiv_id":"2608.11663","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"Zero-OVCD generates change pseudo-labels from SAM3, DINOv3, and SegEarth-OV3, then trains a change detector on them, lifting F1 to 88.65%, 88.85%, and 57.96% on LEVIR-CD, WHU-CD, and S2Looking without target-domain pixel labels.","lead":"Zero-OVCD combines frozen vision foundation models with pseudo-label learning to detect user-specified land-cover changes in bitemporal satellite images without any pixel-level annotations. It reports strong F1 gains on four change detection benchmarks, suggesting a path to cheaper and more flexible change detection in Earth observation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage I threshold selection appears to have been performed on the test splits, so the reported annotation-free gains may partly reflect test-set hyperparameter tuning.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the sensitivity analyses in Sec. IV-E-2 that determine the default thresholds are reported with values identical to the Stage I test results in Table I, yet the evaluation split is not stated. I agree this is the least secure condition for the central claim. The claim is not merely that the pipeline can work well after tuning; it is that Zero-OVCD achieves strong performance without target-domain pixel-level annotations. Selecting thresholds using ground-truth-derived test metrics, even aggregate metrics, violates that framing and makes the reported numbers an upper bound rather than a faithful estimate of annotation-free performance. The concern is concrete and testable. It does not, by itself, invalidate the method: the ablation in Table III shows large contributions from MRM, SMFM, and MCCM under fixed settings, and Table VIII shows consistent gains across foundation-model combinations; moreover, the sensitivity ranges for tau_1, tau_2, and tau_delta are narrow, so the headline F1 may not collapse under honest threshold selection. But the coupled MCCM thresholds are chosen at maxima or boundaries of the evaluated curves, and their numerical effect is not reported, so the magnitude of possible inflation is unknown. Secondary issues include the missing value of the angular threshold theta, single-run results without error bars, and the absence of released code; these reinforce the conditional verdict but are not the primary reason for it. Since my concern matches the reader's, the appropriate recommendation is to keep the conditional verdict unchanged.","tokens_in":22345,"tokens_out":6993,"duration_ms":77342,"concrete_test":"Reproduce Stage I using only the official LEVIR-CD and WHU-CD validation splits for threshold selection: grid-search tau_1 in {0.4,...,0.8}, tau_2 in {2,...,10}, tau_delta in {0.1,...,0.5}, and the three coupled MCCM pairs as in Fig. 10, choosing defaults by mean validation IoU/F1. Then freeze those defaults, evaluate on the test splits, and compare the resulting F1/IoU with Table I. If the validation-selected defaults differ from the paper's defaults, or if the test F1 drops by more than one point on either dataset, the reported gains are partly a test-set tuning artifact and the annotation-free claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Zero-OVCD achieves its results without target-domain pixel-level annotations. The sensitivity analyses in Sec. IV-E-2, which fix the default thresholds, report F1/IoU values on LEVIR-CD and WHU-CD that are exactly the Stage I test values from Table I: for tau_1=0.6, LEVIR-CD gives 86.25/75.82 and WHU-CD gives 85.82/75.17, and the same numbers reappear for tau_2=4 and tau_delta=0.1. The paper never states which split was used in these analyses. If they were run on the official test splits, then tau_1, tau_2, tau_delta, and the coupled MCCM pairs (tau_l/eta_l, tau_s/eta_s, tau_h/tau_o) were selected by maximizing metrics computed with ground-truth labels of exactly the images on which the headline F1s are reported. That makes the reported comparison a fitted evaluation rather than a prediction under the announced annotation-free protocol. This is not a purely cosmetic issue: tau_s/eta_s is set to the highest value of a monotone trend, and tau_h/tau_o=(0.75,0.25) is chosen after a marked IoU decrease on either side, so the tuning decisions are not inert. The sensitivity ranges for tau_1, tau_2, and tau_delta are narrow, so tuning alone may not erase the large gap over baselines; nevertheless, the magnitude of the inflation from the coupled MCCM thresholds is not reported numerically, and the current text does not rule out a material overstatement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Zero-OVCD, a two-stage framework for open-vocabulary change detection (OVCD) that requires no target-domain pixel-level annotations. Stage I generates change pseudo-labels by combining three training-free modules: mask refinement (MRM), multiscale margin-based semantic verification (SMFM), and response-guided correction/completion (MCCM), built on SAM3, DINOv3, and SegEarth-OV3. Stage II trains a ChangerEx detector on these pseudo-labels with checkpoint voting and high-agreement sample selection to mitigate residual noise. The authors report Stage I F1 scores of 86.25%, 85.82%, and 50.48% on LEVIR-CD, WHU-CD, and S2Looking, and Stage II scores of 88.65%, 88.85%, and 57.96% respectively; on SECOND, macro-average F1 rises from 47.91% to 50.92%. The paper claims these results demonstrate an effective annotation-free OVCD solution.","tokens_in":22714,"tokens_out":2163,"duration_ms":22362,"significance":"If validated, the work would be a meaningful step for OVCD: it shows that complementary frozen foundation models can generate pseudo-labels strong enough to train a lightweight task-specific detector, and the two-phase noise-aware training idea is reasonable and clearly ablated. The paper's strengths include a detailed algorithmic description with equations, careful ablations of each Stage I module, sensitivity tables for several thresholds, a comparison across foundation-model combinations, an accuracy-efficiency analysis, and a plan to release code. However, the central quantitative claim currently rests on threshold-selection analyses whose evaluation split is not stated, which is a load-bearing omission; the headline numbers may partly reflect test-set hyperparameter tuning rather than a purely annotation-free protocol. The reported gains over prior training-free methods are large, so the issue must be resolved before the results can be taken at face value.","major_comments":[{"comment":"The sensitivity analyses that set the default thresholds report F1/IoU values on LEVIR-CD and WHU-CD identical to the Stage I test results in Table I (e.g., τ1=0.6 gives 86.25/75.82 on LEVIR-CD and 85.82/75.17 on WHU-CD; the same values reappear for τ2=4 and τδ=0.1). The paper does not state whether these analyses used the validation splits or the test splits. If they used the official test splits, then τ1, τ2, τδ, and the coupled MCCM thresholds were selected by maximizing metrics computed from ground-truth labels of exactly the images on which the headline F1 scores are reported. This would make the reported numbers a fitted evaluation rather than a prediction under the announced annotation-free protocol. The paper must either explicitly state that the sensitivity analyses were restricted to validation splits or re-run the threshold selection (and the resulting final evaluations) on a proper validation split, and report the magnitude of any performance change.","section":"Sec. IV-E-2, Tables IV-VI and Fig. 10"},{"comment":"The angular threshold θ in Eq. (7) is a core parameter of the cross-temporal change proposal construction, but its value is never reported in the implementation details or in any sensitivity analysis. Since θ controls which candidate masks are retained as change proposals, it is a free tuning degree of freedom whose omitted value prevents reproducibility and leaves open the possibility of another test-set-tuned choice. The authors should report the value of θ and explain how it was set, preferably with a sensitivity analysis on a validation split.","section":"Sec. III-A-2 and Sec. IV-B-1"},{"comment":"The coupled-threshold sensitivity analysis reports only the mean changed-class IoU over LEVIR-CD and WHU-CD, computed from what appear to be test splits. The text states that τs/ηs is set to (0.9,0.6) because it is the highest point of a monotone increasing trend and τh/τo to (0.75,0.25) after a marked IoU decrease on either side; these choices are not inert and their effect on the reported gains is not quantified. Even if the individual ranges for τ1, τ2, and τδ are narrow, the MCCM thresholds are coupled and may materially affect the final pseudo-labels. The authors should provide validation-split sensitivity curves and report the F1/IoU differences between the chosen defaults and nearby alternatives.","section":"Sec. IV-E-2, Fig. 10"}],"minor_comments":[{"comment":"The default parameters list mentions τc=0.95, scales {0.8,1.0,1.2}, and γ=500, but not θ; the value of θ should be added here for completeness.","section":"Sec. IV-B-1"},{"comment":"The voting rule assigns a 2:2 tie to the changed state for K=3, which is a deliberate asymmetry in 'equal-weight' voting; this should be stated explicitly as a design choice and its effect analyzed or at least commented on.","section":"Sec. III-B, Eq. (21)"},{"comment":"The sensitivity tables report 'Average' over LEVIR-CD and WHU-CD only, while Fig. 10 uses mean IoU over the same two datasets; this choice should be clarified, especially because S2Looking has a substantially different change distribution and is excluded from the threshold-setting analyses.","section":"Sec. IV-E-2"},{"comment":"The caption says asterisks indicate default settings, but the markers are not visible in the black-and-white rendering of the figure; please ensure the markers and their meaning are legible in the camera-ready version.","section":"Fig. 10"},{"comment":"The ablation baseline is described as 'without MRM, SMFM, or MCCM,' but the composition of that baseline (e.g., which candidates are retained and how semantic verification is performed) is not fully specified; a sentence giving the exact baseline protocol would improve reproducibility.","section":"Sec. IV-E-1, Table III"}],"recommendation":"major_revision","confidential_remarks":"The identical numerical values between the sensitivity tables (Tables IV-VI) and the Stage I test results (Table I) strongly suggest that the threshold-selection analyses were run on the official test splits. If so, the central claim of being annotation-free is undermined because ground-truth test labels were used indirectly to choose defaults. I would ask the authors to clarify the split usage and to re-run the affected analyses on a true validation split; the paper may still have value after that correction, but the headline numbers must be repaired."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a substantial, well-engineered paper with a genuinely new bridging idea, but the stress-test note is right. The sensitivity tables in Sec. IV-E-2 report exact Stage I test numbers for LEVIR-CD and WHU-CD (86.25/75.82 and 85.82/75.17) without stating which split was used. Those are the same numbers as Table I. That is too much of a coincidence; the threshold selection was almost certainly done on the test splits. That makes the reported 'annotation-free' results fitted rather than predicted.\n\nWhat is genuinely new: existing training-free OVCD methods stop at corrected masks; Zero-OVCD closes the loop by using those masks as pseudo-labels to train a task-specific change detector, with checkpoint voting and high-agreement sample selection to fight label noise. That is a real contribution, and the ablation study shows each module adds something. The writing is clear, the equations are mostly precise, and the comparison includes the current SOTA methods. The paper also admits its own limitations (fixed thresholds, per-category training cost, shared errors), which I appreciate.\n\nThe soft spots: the split disclosure is the load-bearing one. If the thresholds were chosen on test data, the headline F1 gains over CoRegOVCD and MemOVCD are overstated by some unknown amount. The sensitivity ranges for tau_1, tau_2, tau_delta are narrow, so those may not matter much, but the MCCM coupled thresholds (tau_s/eta_s, tau_h/tau_o) show clear peaks (tau_s/eta_s at the highest evaluated value, tau_h/tau_o after a marked IoU drop), meaning those choices are not inert. The paper never reports the angular threshold θ at all, so the change-proposal criterion in Eq. (7) is not reproducible. No code is available yet, and results are single-run without error bars. These are fixable, but they matter.\n\nBottom line: the idea deserves a serious referee. I would send it out for review, but the authors need to rerun all threshold selection on validation-only splits, report θ and all hyperparameters, release code, and ideally add variance across runs. Without those, the numbers in the abstract should not be trusted.\n\nWho this is for: anyone working on open-vocabulary change detection or weak supervision with foundation models. The reading group could get a good discussion out of the test-tuning issue alone.","headline":"The two-stage pseudo-label learning idea is real, but the headline F1 numbers are likely tuned on the test splits, so treat the gains as unverified until the authors rerun their sensitivity analysis on validation data.","tokens_in":23237,"tokens_out":2499,"would_cite":false,"duration_ms":24881,"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":"The paper claims that a text-prompt-only, two-stage pipeline — frozen foundation models producing pseudo-labels followed by a noise-aware-trained detector — outperforms prior open-vocabulary change detection on four benchmarks without any…","keywords":["open-vocabulary change detection","pseudo-label learning","training-free inference","vision foundation models","remote sensing","noise-aware training","mask refinement"],"falsifier":"Re-run the sensitivity analysis of Sec. IV-E-2 on the LEVIR-CD and WHU-CD validation splits only, freeze the resulting default thresholds, and then measure test-set F1; if the score drops substantially below the reported 86.25% and 85.82%, the headline numbers depended on test-set information. Alternatively, apply the published thresholds unchanged to a new open-vocabulary change-detection benchmark with different sensors or cities and check whether Stage I still beats the prior best training-free baseline there.","tokens_in":22121,"feed_emoji":"🛰️","tokens_out":17107,"duration_ms":130945,"temperature":0.7,"pith_summary":"The paper claims that open-vocabulary change detection — finding user-specified land-cover changes between two images of the same area — can be solved without any pixel-level labels from the target domain, using only frozen vision foundation models and a text prompt. Stage I turns the frozen models' raw outputs into high-quality change pseudo-labels through three refinement steps: merging automatic and text-guided candidate masks, fusing multiscale semantic similarity scores with margin-based filtering, and correcting or completing masks using cross-temporal response maps. Stage II trains a lightweight change detector on those pseudo-labels while checkpoint voting and high-agreement sample selection suppress the residual label noise. On three building-change benchmarks the pipeline reports F1 scores of 86.25%, 85.82%, and 50.48% at Stage I and 88.65%, 88.85%, and 57.96% at Stage II, plus a macro-average F1 of 50.92% across six categories on the SECOND dataset. If these results hold, open-vocabulary change detection becomes a prompt-driven service: a user types a category and a detector for that category is trained without manual annotation.","feed_headline":"Text prompts alone reach 88.6% F1 in change detection","feed_subtitle":"Frozen vision models hand the detector clean pseudo-labels, beating prior open-vocabulary methods on four benchmarks","key_machinery":"The central mechanism is the two-stage pseudo-label loop. Stage I chains three corrective modules over frozen foundation models. MRM (Mask Refinement Module) fuses class-agnostic masks from SAM3 with text-guided semantic masks, keeping automatic masks whose non-overlap ratio $r_i \\geq \\tau_1$ and dropping those whose containment count $n_i \\geq \\tau_2$; SMFM (Similarity-based Multiscale Fusion Module) averages SegEarth-OV3 category-similarity maps over scales $\\{0.8, 1.0, 1.2\\}$, computes mask-level foreground–background margins $\\delta_i^t$, reassigns low-margin predictions to background, and keeps only automatic-source proposals whose calibrated categories differ across the two time points; MCCM (Mask Correction and Completion Module) filters proposals by target-response coverage under area-adaptive thresholds ($\\tau_l/\\eta_l$ for large masks, $\\tau_s/\\eta_s$ for small ones, split at $\\gamma=500$ pixels) and adds connected components of the cross-temporal high/low-response discrepancy region. Stage II trains ChangerEx, a lightweight bitemporal exchange network, with a composite loss (symmetric cross-entropy plus Lovász-Softmax plus Dice), retains the last three record-improving checkpoints, fuses their predictions with the original pseudo-labels by equal-weight pixel voting (a $2{:}2$ tie counts as changed), and selects high-agreement samples for a reinitialized training run. The load-bearing identity is that pseudo-labels refined by complementary evidence from several frozen models are a sufficient supervision signal for a task-specific detector.","core_discovery":"On its own terms, the paper establishes that training-free foundation-model inference and pseudo-label-supervised detector learning can be bridged into a working open-vocabulary change detection system with no target-domain pixel annotations. Its central discovery is that a progressive refinement chain — MRM merging SAM3 automatic masks with text-guided semantic masks, SMFM aggregating multiscale category-similarity maps under a foreground–background margin filter, and MCCM exploiting bitemporal target-response discrepancies to delete false candidates and recover missed regions — converts noisy foundation-model outputs into pseudo-labels accurate enough to supervise a task-specific change detector. The noise-aware training strategy (record-checkpoint voting, tie-as-changed fusion, and F1-agreement sample selection) then lets the detector exceed the quality of the labels themselves. The paper reports that this two-stage design outperforms existing open-vocabulary change detection methods on LEVIR-CD, WHU-CD, S2Looking, and SECOND, with all foundation models kept frozen and the only user input being the target-category text prompt. The conclusion also flags, as open limitations, that fixed thresholds may not generalize across imaging conditions, that per-category Stage II adaptation raises cost, and that agreement-based sample selection may preserve shared prediction errors.","pith_inferences":["A direct portability test would freeze the published thresholds and run Stage I on an unseen pair of datasets without any re-tuning; the paper's own cross-benchmark spread (88.65% F1 on LEVIR-CD versus 57.96% on S2Looking at Stage II) hints at how much of the gain may depend on dataset-specific threshold settings.","The tie-as-changed voting rule (a 2:2 split among four voters counts as changed) embeds an implicit recall bias; an inexpensive variant would make the tie-break tunable or weight each checkpoint's vote by its proxy-validation F1 instead of voting equally.","Since the text prompt and its background contrast are the only semantic inputs, the method's accuracy likely depends on prompt wording; a prompt-sensitivity study with non-expert phrasings would show whether the reported numbers survive casual usage.","The per-category gains on SECOND — largest for building and low vegetation, smallest for playground and water — suggest that the frozen models' zero-shot semantic alignment is category-dependent; measuring per-category calibration and choosing prompts or fusion weights per category could reveal where the pipeline's ceiling sits."],"forward_implications":["Open-vocabulary change detection becomes a prompt-driven procedure: with a category typed in text, a user obtains pseudo-labels and a trained detector for that category, with no pixel-level annotation effort in the target domain.","Stage II consistently improves over Stage I (for example 86.25 to 88.65 F1 on LEVIR-CD), implying that noise-aware training — voting across checkpoints and keeping high-agreement samples — lets a detector extract signal from imperfect pseudo-labels rather than merely memorizing their errors.","The refinement chain improved all four foundation-model combinations tested, suggesting that the MRM–SMFM–MCCM design is a portable layer that can be re-applied as newer frozen segmenters, encoders, and similarity models appear.","The SECOND results show the method is not limited to building change: six land-cover categories are handled one-vs-rest with the same pipeline, and the macro-average F1 rises from 47.91% at Stage I to 50.92% at Stage II.","Because the deployed Stage II detector is lightweight (11.39M parameters and roughly 12 ms per image pair on WHU-CD), the expensive multi-foundation-model Stage I can be amortized into a small detector for repeated use."],"supporting_citations":[{"why":"Defines open-vocabulary change detection, supplies the M-C-I and I-M-C training-free paradigms that Zero-OVCD builds on, and provides the reproduced baselines and the SECOND one-vs-rest protocol it must beat.","marker":"[13]"},{"why":"SAM3, the frozen segmenter whose automatic and text-guided modes generate the candidate masks that MRM refines.","marker":"[31]"},{"why":"DINOv3, the frozen encoder whose bitemporal mask-level features determine change proposals by cosine similarity.","marker":"[33]"},{"why":"SegEarth-OV3, the frozen open-vocabulary segmenter that produces the multiscale category-similarity maps used by SMFM and MCCM.","marker":"[36]"},{"why":"Changer, the source of the ChangerEx detector architecture trained in Stage II.","marker":"[49]"},{"why":"Symmetric cross-entropy loss, the component that makes Stage II training robust to noisy pseudo-labels.","marker":"[51]"},{"why":"LEVIR-CD, one of the three binary building-change benchmarks the headline F1 scores are reported on.","marker":"[54]"},{"why":"WHU-CD, the second binary benchmark, also used in the accuracy–efficiency comparison.","marker":"[55]"},{"why":"S2Looking, the third binary benchmark, where the largest Stage I gain over prior methods is reported.","marker":"[56]"},{"why":"SECOND, the six-category benchmark used for the category-wise one-vs-rest evaluation.","marker":"[57]"}],"fun_headline_variants":["Zero-OVCD bridges frozen models and pseudo-labels for open-vocab change detection","Training-free foundation models hand detectors clean pseudo-labels","No target-domain labels: Zero-OVCD lifts F1 to 88.6% on LEVIR-CD","Two-stage zero-label pipeline outperforms prior open-vocab change detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The performance claims assume the pseudo-label thresholds in Stage I (for example $\\tau_1=0.6$, $\\tau_2=4$, $\\tau_\\delta=0.1$, $\\gamma=500$) were chosen without peeking at test labels: the sensitivity analyses in Sec. IV-E-2 report scores on LEVIR-CD and WHU-CD but never state which split they evaluate on, so if the test labels shaped the defaults, the reported F1 figures overstate what an annotation-free user would achieve on unseen data.","fun_headline_variants_meta":{"raw":{"variants":["Zero-OVCD bridges frozen models and pseudo-labels for open-vocab change detection","Training-free foundation models hand detectors clean pseudo-labels","No target-domain labels: Zero-OVCD lifts F1 to 88.6% on LEVIR-CD","Two-stage zero-label pipeline outperforms prior open-vocab change detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000367,"raw_usage":{"total_tokens":2046,"prompt_tokens":1094,"completion_tokens":952,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":865}},"tokens_in":710,"tokens_out":952,"duration_ms":8977,"temperature":1.0,"reasoning_tokens":865,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:31:47.929939+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the sensitivity analysis of Sec. IV-E-2 on the LEVIR-CD and WHU-CD validation splits only, freeze the resulting default thresholds, and then measure test-set F1; if the score drops substantially below the reported 86.25% and 85.82%, the headline numbers depended on test-set information. Alternatively, apply the published thresholds unchanged to a new open-vocabulary change-detection benchmark with different sensors or cities and check whether Stage I still beats the prior best training-free baseline there.","supporting_citations":[{"cited_title":"DynamicEarth: How far are we from open-vocabulary change detection?","cited_arxiv_id":null,"evidence_quote":"Defines open-vocabulary change detection, supplies the M-C-I and I-M-C training-free paradigms that Zero-OVCD builds on, and provides the reproduced baselines and the SECOND one-vs-rest protocol it must beat."},{"cited_title":"SAM 3: Segment anything with concepts,","cited_arxiv_id":null,"evidence_quote":"SAM3, the frozen segmenter whose automatic and text-guided modes generate the candidate masks that MRM refines."},{"cited_title":"Sim ´eoni, H","cited_arxiv_id":null,"evidence_quote":"DINOv3, the frozen encoder whose bitemporal mask-level features determine change proposals by cosine similarity."},{"cited_title":"Symmetric cross entropy for robust learning with noisy labels,","cited_arxiv_id":null,"evidence_quote":"Symmetric cross-entropy loss, the component that makes Stage II training robust to noisy pseudo-labels."}],"review_version":1}