CogVis: Must Open-Vocabulary Change Detection Perceive the Scene Anew for Every Query?

CogVis: Must Open-Vocabulary Change Detection Perceive the Scene Anew for Every Query? Zijie Wang, Chen Zhong, Wei He∗ Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University Wuhan, China zijie.wang@whu.edu.cn, chen_zhong@whu.edu.cn, weihe1990@whu.edu.cn

arXiv:2608.06150v1 [cs.AI] 6 Aug 2026

Abstract Earth-surface monitoring requires change detection models capable of recognizing arbitrary semantic categories. OpenVocabulary Change Detection (OVCD) addresses this need. However, existing methods often entangle temporal perception, semantic discrimination, and region verification, causing unstable results and redundant computation. Inspired by human visual change perception, we propose CogVis, a cognitive memory-guided framework that reformulates OVCD as a perception-memory-verification paradigm. CogVis first employs a Scene Change Perceptron (SCP) to extract a reusable, category-agnostic change prior from frozen bi-temporal features, thereby decoupling temporal evidence from semantic category decisions. A Semantic Memory Calibrator (SMC) then compensates for category-dependent score shifts by dynamically estimating an image-query-specific decision threshold. Finally, an Adaptive Region Filter (ARF) filters connected candidates using learned semantic, temporal, and structural reliability. Experiments on seven benchmarks spanning semantic change detection, binary change localization, and buildingdamage assessment show that CogVis achieves state-of-theart performance across all evaluated datasets. By sharing scene-level change perception, CogVis further avoids repeating category-agnostic temporal perception across queries and improves inference throughput by 28.50%.

Code — https://github.com/KotlinWang/CogVis

Introduction Earth’s surface is continuously reshaped by natural processes and human activities, producing complex spatiotemporal patterns that must be reliably interpreted for environmental monitoring and informed decision-making (Li et al. 2026a). Change Detection (CD) identifies changes in multitemporal remote sensing imagery and supports applications including land-cover monitoring, disaster assessment, urban planning, and ecological observation (Hong et al. 2024; Wang et al. 2025). However, most CD research assumes that scene distributions and semantic categories are fixed during training. This closed-set assumption limits their applicability in openworld scenarios, where acquisition conditions vary and users may query categories that are absent from the task-specific classifier. ∗

Corresponding author.

(a)

Mask Proposal (M)

Comparator (C)

Regional Reliability (b)

VS

VS

VS

VS

Identifier (I) Building Water

Temporal Evidence

Ground

Semantic Uncertainty

Scene-level Changes Semantic Memory Region Check

VS

VS

Perception

Memory

Verification

Figure 1: Comparison of OVCD paradigms. Existing M-C-I and I-M-C pipelines couple temporal, semantic, and regional decisions, whereas CogVis separates them into shared perception, memory-guided calibration, and verification.

Vision models and Vision-Language Models (VLMs) provide transferable priors for segmentation, representation learning, and text-image alignment. SAM3 (Carion et al. 2025), DINOv3 (Siméoni et al. 2025), and CLIP (Radford et al. 2021) have consequently enabled zero-shot or traininglight remote sensing interpretation. Foundation-model-based change detectors such as AnyChange (Zheng et al. 2024) and UCD-SCM (Tan et al. 2024) improve category-agnostic change localization but primarily determine whether a region has changed, not which text-specified category accounts for the change. Open-Vocabulary Change Detection (OVCD) has emerged to bridge this gap by detecting bi-temporal changes under text-specified concepts. DynamicEarth (Li et al. 2026b) summarizes existing research through the MaskCompare-Identify (M-C-I) and Identify-Mask-Compare (IM-C) pipelines. M-C-I shares category-independent proposals, but downstream recognition inherits their localization errors. Missed candidates cannot be recovered, whereas spurious candidates may still receive plausible labels (Dou et al. 2026; Guo et al. 2026). I-M-C injects category priors before temporal comparison, enabling category-aware reasoning but repeatedly parses the image pair for every query (Zhu et al. 2025; Zhang et al. 2026b; Tang et al. 2026; Kuang et al. 2026). Therefore, both designs couple temporal evidence, semantic uncertainty, and regional reliability within a fixed cascade, as shown in Fig. 1(a).

Unlike the fixed logic of M-C-I or I-M-C, we formulate OVCD as a perception-memory-verification paradigm motivated by human visual change perception. As shown in Fig. 1(b), humans first form a global perception of scenelevel changes, then leverage semantic memory to infer plausible categories, and finally verify each candidate region to reject non-semantic disturbances such as shadows, texture variations, seasonal shifts, and misregistration (Lerebourg, de Lange, and Peelen 2024; Brady, Robinson, and Williams 2024). Based on this observation, we propose CogVis, a cognitive memory-guided OVCD framework that follows a scene-first and memory-calibrated process. CogVis first extracts a reusable category-agnostic temporal prior with a Scene Change Perceptron (SCP). A Semantic Memory Calibrator (SMC) then contrasts each query with nearby semantic confusers and retrieves memory cases to estimate an imagequery-specific decision threshold. Finally, an Adaptive Region Filter (ARF) verifies connected candidates through semantic, temporal, and structural agreement. The main contributions of this work are as follows: • Inspired by human visual change perception, we reformulate OVCD as a perception-memory-verification paradigm that disentangles the decisions coupled within the fixed M-C-I and I-M-C cascades. • We instantiate this paradigm as CogVis, where SCP, SMC, and ARF jointly address semantic-temporalstructural entanglement, image-query score shifts, and unreliable region predictions. • We validate this factorization on seven benchmarks, demonstrating consistent cross-task gains and reduced multi-query temporal computation.

Related Work Foundation Models for Remote Sensing Foundation models are increasingly replacing task-specific representation learning in remote sensing, reflecting a broader shift toward universal large-scale remote sensing models (Zhang et al. 2026a). CLIP aligns imagery with open-vocabulary concepts (Radford et al. 2021; Liu et al. 2024), DINO-style encoders provide spatially coherent dense features (Caron et al. 2021; Oquab et al. 2024), and SAMstyle models generate promptable masks (Kirillov et al. 2023; Carion et al. 2025). Methods such as SegEarth-OV, ReSegCLIP, and SeeCo further adapt these priors to remote-sensing challenges including scale variation, weak boundaries, and domain-specific semantics (Li et al. 2025b; Heidarianbaei et al. 2026; Wang et al. 2026). However, these advances primarily target single-image interpretation. Change understanding additionally requires temporally comparable evidence while preserving open-vocabulary semantics. Therefore, CogVis retains the frozen semantic pathway and learns only lightweight adapters for temporal perception, calibration, and verification.

Open-Vocabulary Change Detection OVCD extends CD from fixed categories to text-specified semantic changes. DynamicEarth formalizes the dominant

M-C-I and I-M-C paradigms (Li et al. 2026b). AdaptOVCD aligns heterogeneous model outputs through adaptive fusion (Dou et al. 2026), OpenDPR retrieves visual prototypes for open-category recognition (Guo et al. 2026), Seg2Change transfers open-vocabulary segmentation to CD (Su et al. 2026), and OmniOVCD addresses semantic and instance mismatches (Zhang et al. 2026b). Although these methods improve individual stages, they do not explicitly separate query-shared temporal perception, query-specific boundary calibration, and region-level verification. CogVis instead organizes these decisions according to their scope rather than embedding them in a fixed processing cascade.

Memory for Change Reasoning In conventional CD, memory typically stores temporal states or feature prototypes. MS-Former maintains changed and unchanged prototypes, whereas ChangeTitans employs neural memory for long-range temporal modeling (Li et al. 2024; Yang et al. 2025). MemOVCD further uses memory for crosstemporal semantic propagation and global-local rectification (Kuang et al. 2026). CogVis adopts a fundamentally different role for memory. Each entry associates a continuous scenetemporal semantic response state with a calibrated threshold. Therefore, retrieval transfers an operating point rather than a feature prototype, which the Score Adapter further adjusts for the current image-query pair. This design directly compensates for scene- and category-dependent score shifts without modifying the semantic response map.

Methodology Problem Formulation: Given bi-temporal remote sensing images I 1 , I 2 ∈ RH×W ×3 and the open vocabulary C = {c1 , . . . , cK }, OVCD predicts a change mask for each query  M = Mc ∈ {0, 1}H×W | c ∈ C . (1) The task contains three decisions with different scopes. Temporal evidence is shared across queries, semantic confidence depends on the image-query pair, and candidate regions require reliability verification. CogVis addresses them with SCP, SMC, and ARF, respectively. Framework Overview: As shown in Fig. 2, a frozen visual encoder extracts bi-temporal features. SCP compares the image pair once to produce a reusable change prior B. For each query c, SMC decodes the original frozen features, constructs a semantic-change response Rc , and estimates an adaptive threshold τc . ARF then verifies connected candidates to obtain Mc . The SCP residual is used only for B and never alters semantic decoding.

Scene Change Perceptron SCP extracts query-shared temporal evidence once per image t pair. Let {Ftl }L l=1 = E(I ) denote the frozen SAM3 feature pyramid at time t ∈ {1, 2}. We denote the second-level features by Ft = Ft2 and construct
X = Catch F1 , F2 , |F2 − F1 |, F1 ⊙ F2 . (2) b = W↑ ϕ(W↓ X). A Change Adapter maps X to a residual ∆ A temporal gate produces ∆ by capping its root-meansquare magnitude at 25% of the frozen-feature magnitude.

CogVis Workflow 𝑰𝟏

SCP

SMC

𝑩

𝐅!"

Concat 𝐅!" , 𝐅!#

Encoder

Inputs

𝐅#$

𝑭𝟐𝒍 − 𝑭𝟏𝒍

𝑰" Encoder 𝐅#" Frozen

Abs Diff

Product

𝑹𝒄

Score… Adapter

Decoder Change Adapter

𝐅!# Building Tree u Water u Ground u Playground u …

Top-K 𝝉 𝒄 Retrieval

u u

𝚫

𝐅!" ⊙ 𝐅!#

𝑹𝒄

Perception

CLIP

ARF Area

Score

Margin

Prior

…

)% = 𝐌

𝟏 𝐑 ! ≥ 𝜏!

𝑩

Memory bank Memory

Outputs

Gate Adapter 𝒖𝒋 𝒖𝒋

≥

Verification

𝐌%

𝐌%

𝝉𝒓 Trainable

Figure 2: Overview of CogVis. SCP extracts a query-shared change prior, SMC calibrates query-specific responses with memory, and ARF verifies candidate regions. The change head then predicts

B = σ Up Hchg (X, ∆) .

(3)

The resulting category-agnostic prior is shared by all queries, while the original feature pyramid is preserved for semantic decoding.

Semantic Memory Calibrator Open-vocabulary responses vary across scenes and categories, making a fixed decision boundary unreliable. Therefore, SMC constructs a query-specific semantic-change response and calibrates its threshold from memory. Semantic-change Response: For query c, the frozen decoder processes prompt set Pc using the original temporal features. Let Qtc denote the aggregated target response at time t, and Ctc the strongest response among CLIP-selected semantic neighbors. Prompt competition suppresses ambiguous activation as ρ  Qtc e t = Qt ⊙ , t ∈ {1, 2}, (4) Q c c Qtc + Ctc + ϵ where ρ controls competition strength. The competitionadjusted responses produce four cues. Dc captures semantic transitions, Gc enforces support from the change prior B, Sc suppresses persistent semantics in stable regions, and pc measures the change in query presence. Their agreement defines Rc = Π[0,1] (Dc ⊙ Gc ⊙ Sc pc ) , (5) The resulting response retains query-specific transitions supported by temporal evidence. Memory Construction and Retrieval: Each source imagequery pair stores a normalized 1556-dimensional key and its oracle threshold. The key concatenates scene context, temporal discrepancy, and query embeddings in R512 with 20 statistics describing Rc and its agreement with B. The oracle threshold maximizes IoU between the thresholded response and the query-specific source mask over a predefined grid. At inference, cosine similarity retrieves the top Kmem = 16 memory entries. Their thresholds form a temperatureweighted anchor X exp(si /Tm ) , τ̄c = πic τi⋆ , (6) πic = P j∈Kc exp(sj /Tm ) i∈Kc

where sci = k⊤ c ki and Tm = 0.07. A Score Adapter predicts a bounded correction ∆τc from the response statistics, retrieved anchor, and similarity confidence. Combined with the label-free threshold τc0 = Tadp (Rc ), the boundary is
τc = Π[τmin ,τmax ] max{τ̄c + ∆τc , 0.75τc0 } . (7) Here, Π[a,b] (·) denotes element-wise clipping to [a, b]. The retrieved anchor transfers a suitable operating point, while the learned correction adapts it to the current image-query fc. pair. Thresholding Rc at τc yields the candidate mask M

Adaptive Region Filter Pixelwise thresholding may retain fragmented or weak responses. Therefore, ARF verifies connected regions bef c ) denote the candidate fore prediction. Let Ωc = CCL(M components. The CCL denotes connected-component labeling. After deterministic component and area handling, each ωj ∈ Ωc is encoded by a 19-dimensional descriptor φj covering geometry, semantic confidence, temporal support, calibrated margin, and semantic-temporal agreement. The Gate Adapter predicts its reliability as uj = σ(Agate (φj )). Let aj = |ωj |/(HW ) denote the area ratio of component ωj . We set the large-region protection threshold to ηa = 0.04 and the Gate threshold to τr = 0.60. ARF retains a component if it satisfies either criterion: [ Mc = {ωj ∈ Ωc | aj ≥ ηa ∨ uj ≥ τr } . (8) This rule preserves large components directly, while smaller candidates must pass learned semantic-temporal-structural verification. ARF removes unreliable proposals but cannot fc. recover regions missing from M

Adapter Training The visual encoder, semantic decoder, and CLIP encoders remain frozen. The Change, Score, and Gate Adapters are optimized sequentially as separate checkpoints. The Change Adapter is trained with the binary-change objective Lchg , comprising equally weighted balanced BCEwith-logits and soft Dice losses, to produce the prior in Eq. 3. Subsequently, the source responses are generated, and the

Method

Venue

Semantic CD (%)

Binary CD (%)

Disaster (%)

SECOND SCSCD CLCD DSIFN LEVIR-CD WHU-CD Category-agnostic CD PCA-KMeans GRSL09 CVA RSE94 DCVA TGRS19 UCD-SCM IGARSS24 AnyChange NeurIPS24 Inst-CEG TGRS25 Cascaded OVCD DynamicEarth-MCI AAAI26 DynamicEarth-IMC AAAI26 UniVCD arXiv25 SAM3-CD arXiv25 AdaptOVCD arXiv26 OpenDPR CVPR26 OmniOVCD arXiv26 CoRegOVCD arXiv26 Seg2Change arXiv26 Memory-augmented OVCD MemOVCD arXiv26 CogVis (Ours) – Margin –

#P ↓

xBD

− − − 8.40 11.84 10.46

− − − 6.51 9.32 3.68

10.14 7.17 11.62 13.19 19.02 3.50

21.57 16.20 23.85 25.10 24.37 18.91

5.24 4.86 7.44 19.30 19.53 46.30

7.72 3.72 11.45 19.14 16.37 45.49

− − − 19.04 19.63 −

− − 10 219 641.09 −

23.58 13.72 20.53 16.12 28.05 28.55 27.05 31.67 29.08

17.97 5.16 − − − − − − 23.22

13.52 8.09 − − − − − − 31.48

37.32 15.22 − − 42.32 − 24.50 47.55 41.40

30.23 53.50 56.70 40.00 51.52 44.80 67.20 70.95 64.91

40.20 61.09 66.00 36.30 61.99 54.30 66.50 69.70 75.72

24.51 − − − 22.07 19.58 18.92 − 17.94

1087 584 − − 1393.78 727.67 840.51 − 931.01

− 32.48 +0.81

− 25.36 +2.14

− 33.67 +2.19

37.80 48.50 +0.95

72.50 72.55 +0.05

− 75.86 +0.14

− 32.68 +8.17

− 990.29 −

Table 1: Comparison on seven benchmarks. We report mIoU for semantic CD and xBD and IoU for binary CD. Bold and underlined indicate the best and second-best results, respectively; “−” denotes an unreported result. “Margin” is the gain over the strongest prior result, and #P denotes total parameters (M). calibration memory is constructed offline. For source entry (−i) i, leave-one-out retrieval yields τ̄i , and the Score Adapter regresses the residual to the oracle threshold using SmoothL1 loss: N   1 X (−i) Lscore = ℓsl1 ∆τi , τi⋆ − τ̄i . (9) N i=1 Finally, connected candidates are labeled by the sourceoverlap rule, and the Gate Adapter is trained with binary cross-entropy: Lgate =

NΩ 1 X ℓbce (uj , yj ), NΩ j=1

(10)

where yj ∈ {0, 1} is the target reliability of ωj . This staged procedure preserves the role of each adapter and avoids coupling the three decision stages during optimization.

Experiments and Analysis Experimental Settings Benchmarks: We evaluate CogVis on seven public benchmarks covering semantic CD, including SECOND (Yang et al. 2022) and SCSCD (Tan et al. 2025); binary CD, including CLCD (Liu et al. 2022), DSIFN (Zhang et al. 2020), LEVIR-CD (Chen and Shi 2020b), and WHU-CD (Ji, Wei, and Lu 2019); and building-damage assessment on xBD (Gupta et al. 2019). CA-CDD (Su et al. 2026) is used exclusively for adapter optimization and memory construction and does not overlap with any test set.

Metrics: We report mIoU as the primary metric for semantic CD and building-damage assessment, and IoU for binary CD. Efficiency is measured by latency, frames per second (FPS), peak GPU memory, and parameter count. Query-scaling experiments use 512 × 512 inputs and are conducted on an RTX 3090 GPU. Implementation: CogVis uses frozen SAM3 (Carion et al. 2025) visual features and CLIP ViT-B/16 (Radford et al. 2021) embeddings for text encoding and retrieval. The Change, Score, and Gate Adapters are optimized sequentially on CA-CDD for 100 epochs with a batch size of 16. The top 16 memory-bank entries are retrieved with a temperature of 0.07. The ARF gating and large-region thresholds are set to 0.60 and 0.04, respectively.

Comparison with State-of-the-Art Methods Quantitative Comparison: CogVis achieves the best metric on all seven benchmarks, surpassing the strongest competitor by 0.05–8.17 percentage points, as shown in Table 1. It obtains 32.48% and 25.36% mIoU on SECOND and SCSCD, improving the previous best results by 0.81 and 2.14 points. For binary CD, CogVis reaches 33.67%, 48.50%, 72.55%, and 75.86% IoU on CLCD, DSIFN, LEVIR-CD, and WHUCD, corresponding to gains of 2.19, 0.95, 0.05, and 0.14 points. The larger gains on CLCD and DSIFN are consistent with the framework’s ability to suppress appearanceinduced temporal nuisances. The smaller margins on the cleaner LEVIR-CD and WHU-CD datasets reflect the already strong performance of existing methods. On xBD, CogVis improves mIoU from 24.51% to 32.68%, producing

T2

GT

UCD-SCM

AnyChange DynamicEarth-MCI AdaptOVCD

OpenDPR

OmniOVCD

Seg2Change

CogVis (Ours)

xBD

SECOND

SCSCD

DSIFN

CLCD

LEVIR-CD

WHU-CD

T1

Change / Building

Water

Vegetation / Tree

Farmland / Grass

Bareland / Ground

Structure

Road

Playground

No-Damage

Minor-Damage

Major-Damage

Destroyed

Figure 3: Qualitative comparison across seven benchmarks covering semantic CD, binary CD, and building-damage assessment. Ground truth and predictions follow the class legend. the largest gain of 8.17 points. These consistent improvements across all three tasks demonstrate the transferability of the perception-memory-verification design. Qualitative Comparison: As shown in Fig. 3, CogVis produces more coherent change regions and suppresses appearance-induced false responses across the seven benchmarks. It better separates queried categories on semantic and damage-assessment benchmarks while improving change localization on binary benchmarks. Figure 4 further presents representative error maps under shadows, radiometric or textural variation, seasonal shifts, and residual misregistration. CogVis often yields sharper boundaries and predictions that more closely align with the annotations. These examples illustrate its robustness to non-semantic perturbations through change-prior modeling, query-specific calibration, and region-level reliability filtering. Efficiency and Scalability: CogVis reaches 0.230 FPS with 4.82 GiB peak memory, providing 28.50% higher throughput than the next-fastest method in Fig. 5. Since CogVis is not the smallest model, this gain arises from shared temporal computation rather than parameter reduction. At K = 10, CogVis requires 10.03 s, compared with 12.56– 956.60 s for category-wise baselines. Latency still increases with the number of queries because SMC and ARF remain

query-specific, but one-time temporal perception substantially reduces the per-query cost.

Ablation Studies Component Ablation: The full model outperforms Vanilla by 13.12 mIoU on SECOND, 21.57 IoU on CLCD, and 4.25 mIoU on xBD, as shown in Table 2. Among individual components, SCP delivers the largest gains on SECOND and CLCD with improvements of 2.95 and 8.13 points, whereas SMC contributes the largest gain on xBD at 1.85 points. These results align with their respective roles in temporal stabilization and category-dependent calibration. Combining SCP and SMC produces the strongest two-component variant on all three benchmarks, confirming their complementarity. However, ARF depends strongly on the quality of upstream predictions. Used alone, it slightly reduces performance on SECOND while improving CLCD and xBD by 3.22 and 0.17 points. Combined only with SMC, it slightly degrades performance on SECOND and xBD but improves CLCD by 3.25 points, indicating that region verification remains unstable without a reliable temporal prior. With both SCP and SMC enabled, ARF consistently improves SECOND, CLCD, and xBD by 4.44, 4.76, and 1.82 points, respectively. These findings establish ARF as a final verification stage for

T1

T2

OpenDPR

OmniOVCD

Seg2Change

CogVis (Ours)

CA-CDD

Shadow

Variant

Radiometric

I II III Full

SECOND CLCD

Adapters Memory ✓ × × × × – ✓ ✓

mIoU ↑ 32.10 31.42 23.39 32.48

xBD

IoU ↑ mIoU ↑ 33.32 32.66 33.17 32.46 25.70 30.86 33.67 32.68

Misregistration

Seasonal

Table 3: Source-domain dependence. “✓” and “×” indicate whether the subset of CA-CDD is retained or excluded from adapter training or memory construction. “–” disables memory entirely. Bold marks the best result in each metric column.

True Positive (TP)

False Positive (FP)

True Negative (TN)

False Negative (FN)

Figure 4: Qualitative comparison across four nuisance conditions: shadows, radiometric or textural variations, seasonal shifts, and residual misregistration. Modules

Variant

SECOND CLCD

SCP SMC ARF Vanilla I II III IV V VI VII

✓ ✓ ✓ ✓ ✓ ✓

✓ ✓ ✓

✓ ✓ ✓

mIoU ↑ 19.36 22.31 20.25 19.30 28.04 25.75 20.17 32.48

xBD

#P ↓

IoU ↑ mIoU ↑ 12.10 28.43 840.51 20.23 28.93 840.66 13.96 30.28 990.14 15.32 28.60 840.52 28.91 30.86 990.29 25.87 29.12 840.66 17.21 29.54 990.14 33.67 32.68 990.29

Table 2: Component ablation. Bold marks the best result in each metric column. #P denotes total parameters (M). temporally stabilized and semantically calibrated candidates rather than a generic post-processing module. Modular Ablation: Figure 6 shows that SCP increases ROC AUC from 0.79 with raw RGB differences to 0.96 with the learned change prior. Figure 7 visualizes this progression: retrieval provides a suitable threshold anchor, and samplespecific correction suppresses residual false positives. Figure 8 further shows consistent degradation when semantic, temporal, or geometric evidence is removed from ARF. Together, these results support the distinct and complementary roles of temporal perception, semantic calibration, and region verification. Memory-Bank Ablation: Table 3 tests whether CogVis’s SECOND gains depend on the CA-CDD subset. Memory-only exclusion reduces SECOND/CLCD/xBD by 0.38/0.35/0.02 points, while excluding the subset from adapter optimization and memory construction reduces them by 1.06/0.50/0.22. These small, cross-benchmark decreases are more consistent with reduced source diversity and training volume than with a SECOND-specific advantage. This interpretation matches the adapter objectives, which target category-agnostic temporal evidence, residual threshold correction, and region reliability rather than dataset-specific

Memory SECOND CLCD xBD Efficiency Cal. bank mIoU ↑ IoU ↑ mIoU ↑ Mem. ↓ Lat. ↓ Fix. 26.35 25.88 31.08 0% 0.0 0 Rft. – – – 10%

Fix. Rft.

32.17 32.00

33.49 33.18

32.61 32.64

52.6

0.0012

25%

Fix. Rft.

32.21 32.11

33.18 32.91

32.65 32.66

131.4

0.0030

Fix. Rft. Fix. Rft.

32.27 32.29 32.48 32.48

33.26 32.97 33.67 33.67

32.67 32.66 32.68 32.68

262.6

0.0097

525.2

0.0113

50% 100%

Table 4: Memory-bank scale analysis. “Cal." denotes calibration; “Fix. and Rft." denote frozen and refitted calibration, respectively, while Mem. (MiB) and Lat. (ms/query) report storage and retrieval latency. Bold marks the best accuracy in each metric column.

semantics. Disabling the remaining cross-source memory causes further losses of 8.03, 7.47, and 1.60 points, indicating that the gains derive from transferable responsethreshold knowledge rather than same-dataset memorization. Table 4 shows that a 10% bank retains 94.9%/97.7%/95.6% of full-memory gains on SECOND/CLCD/xBD while reducing storage from 525.2 to 52.6 MiB. At 25% or 50%, changes are marginal, and the small frozen/refitted gap indicates that saturation is not driven by scale-specific retraining, although the full bank remains best overall. Thus, SMC requires representative coverage of the response-threshold space rather than exhaustive source storage, making semantic memory a compact, transferable calibration prior.

Sensitivity Analysis Table 5 shows that CogVis is robust to memory-retrieval settings: varying Kmem from 8 to 32 or Tm from 0.05 to 0.10 causes only marginal changes across all benchmarks. Consistent with the memory-bank ablation, this stability indicates that SMC transfers representative response-threshold relationships without relying on a specific retrieval neighborhood. Region-level parameters have larger effects, particularly on CLCD: an overly permissive τr retains unreliable components, whereas an overly strict τr removes valid

35

0.30

(a) Best region

SCSCD SECOND xBD

CogVis Seg2Change OmniOVCD

OpenDPR AdaptOVCD DynamicEarth-MCI

12

(b)

AnyChange UCD-SCM

Mean FPS Peak memory (GiB)

1,000

(c)

10.32

0.25

10

0.230

30

20

8

0.179 6.40 5.78

5.61

0.15

6 4.82

0.112 3.99

0.10

15

4

Inference time (s)

25

Mean FPS

mIoU (%)

8.23

0.20

100

10

0.066

0.05

10

2

0.036

UCD-SCM AnyChange

0.69 0.006

Poor region

3

5

10

20

30

50

Inference time (s)

100

200 300

500

0.009

0.00

0 UCDSCM

Any Change

Adapt OVCD

Dynamic Earth-MCI

DynamicEarth-MCI AdaptOVCD

OpenDPR OmniOVCD

5

7

CogVis Seg2Change

1

0.003

Open DPR

Omni OVCD

Seg2 Change

1

CogVis

2

3

4

K

6

8

9

10

Figure 5: Accuracy-efficiency and query-scaling analysis on RTX 3090 GPU with 512 × 512 inputs. (a) Inference time versus mIoU. (b) Mean throughput and peak memory. (c) Latency as the number of queried categories increases. T1

T2

Dataset ROC

1.0

32.20

32.30

mIoU (%)

RGB RMS difference

SCP shared prior B

True positive rate

0.8

32.24

32.40

32.50

w/o Semantic

w/o Geometry

w/o Temporal

Full ARF

0.6

32.25 0.4

32.23

0.2 RGB RMS difference AUC 0.789 SCP shared prior B AUC 0.955

32.48

0.0

0.0

0.2

0.4

0.6

0.8

1.0

False positive rate

Figure 6: SCP analysis. The learned change prior suppresses appearance-only differences and improves ROC AUC from 0.79 to 0.96. GT

Base threshold

Memory retrieval

Score Adapter

Class-normalized response margin (max over classes) Below threshold (<0)

Decision boundary (=0)

Figure 8: ARF evidence ablation on SECOND. Removing semantic, temporal, or geometric evidence degrades region verification.

Above threshold (>0)

Figure 7: SMC analysis. Memory retrieval anchors the threshold, which the Score Adapter corrects for each imagequery pair. Gold and magenta contours denote ground truth and false positives.

and ARF, it reuses scene-level change evidence, transfers response-threshold knowledge across domains, and suppresses unreliable regions using complementary semantic, temporal, and structural cues. Experiments on seven benchmarks demonstrate consistent state-of-the-art performance, strong cross-task transferability, and a 28.50% improvement in inference throughput. These results show that openvocabulary change understanding need not reinterpret the scene for every query.

Variant

changes. Lowering ηa also makes large-region protection overly permissive, allowing smaller unreliable components to bypass the learned gate and reducing performance on SECOND and CLCD. The default configuration offers the best overall balance, supporting CogVis’s intended functional separation.

Conclusion This work revisits OVCD from the perspective of decision scope. CogVis separates query-shared temporal perception from query-dependent semantic calibration and regionlevel verification, allowing each source of uncertainty to be addressed at the appropriate stage. Through SCP, SMC,

I II III IV V VI VII Default

Configuration Kmem Tm τr ηa 8 0.07 0.6 0.04 32 0.07 0.6 0.04 16 0.05 0.6 0.04 16 0.10 0.6 0.04 16 0.07 0.5 0.04 16 0.07 0.7 0.04 16 0.07 0.6 0.01 16 0.07 0.6 0.04

SECOND CLCD mIoU ↑ 32.30 32.32 32.29 32.29 32.16 32.17 32.11 32.48

xBD

IoU ↑ mIoU ↑ 33.17 32.67 33.45 32.68 33.32 32.67 33.33 32.67 32.05 32.32 32.33 32.62 32.69 32.58 33.67 32.68

Table 5: Hyperparameter sensitivity. Bold marks the best accuracy in each metric column, and the bold row denotes the default configuration.

References Brady, T. F.; Robinson, M. M.; and Williams, J. R. 2024. Noisy and hierarchical visual memory across timescales. Nature Reviews Psychology, 3(3): 147–163. Carion, N.; Gustafson, L.; Hu, Y.-T.; Debnath, S.; Hu, R.; Suris, D.; Ryali, C.; Alwala, K. V.; Khedr, H.; Huang, A.; Lei, J.; Ma, T.; Guo, B.; Kalla, A.; Marks, M.; Greer, J.; Wang, M.; Sun, P.; Rädle, R.; Afouras, T.; Mavroudi, E.; Xu, K.; Wu, T.-H.; Zhou, Y.; Momeni, L.; Hazra, R.; Ding, S.; Vaze, S.; Porcher, F.; Li, F.; Li, S.; Kamath, A.; Cheng, H. K.; Dollár, P.; Ravi, N.; Saenko, K.; Zhang, P.; and Feichtenhofer, C. 2025. SAM 3: Segment Anything with Concepts. arXiv preprint arXiv:2511.16719. Caron, M.; Touvron, H.; Misra, I.; Jégou, H.; Mairal, J.; Bojanowski, P.; and Joulin, A. 2021. Emerging Properties in Self-Supervised Vision Transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 9650–9660. Chen, H.; and Shi, Z. 2020a. LEVIR-CD: A Large-Scale Remote Sensing Binary Change Detection Dataset. Official dataset and code repository. Accessed July 31, 2026. Chen, H.; and Shi, Z. 2020b. A Spatial-Temporal AttentionBased Method and a New Dataset for Remote Sensing Image Change Detection. Remote Sensing, 12(10): 1662. Dou, M.; Qiu, S.; Hu, M.; Chen, Y.; Ye, H.; Liao, X.; and Sun, Z. 2026. AdaptOVCD: Training-Free Open-Vocabulary Remote Sensing Change Detection via Adaptive Information Fusion. arXiv preprint arXiv:2602.06529. Guo, Q.; Wang, J.; Liu, Y.; and Zhong, Y. 2026. OpenDPR: Open-Vocabulary Change Detection via Vision-Centric Diffusion-Guided Prototype Retrieval for Remote Sensing Imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 20399– 20409. Gupta, R.; Hosfelt, R.; Sajeev, S.; Patel, N.; Goodman, B.; Doshi, J.; Heim, E.; Choset, H.; and Gaston, M. 2019. Creating xBD: A Dataset for Assessing Building Damage from Satellite Imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Heidarianbaei, M.; Dorozynski, M.; Kanyamahanga, H.; Mehltretter, M.; and Rottensteiner, F. 2026. OpenVocabulary Semantic Segmentation in Remote Sensing via Hierarchical Attention Masking and Model Composition. arXiv preprint arXiv:2602.23869. Published in the proceedings of the British Machine Vision Conference Workshops 2025. Hong, D.; Zhang, B.; Li, X.; Li, Y.; Li, C.; Yao, J.; Yokoya, N.; Li, H.; Ghamisi, P.; Jia, X.; Plaza, A.; Gamba, P.; Benediktsson, J. A.; and Chanussot, J. 2024. SpectralGPT: Spectral Remote Sensing Foundation Model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12): 10145– 10161. Ji, S.; Wei, S.; and Lu, M. 2019. Fully Convolutional Networks for Multisource Building Extraction From an Open Aerial and Satellite Imagery Data Set. IEEE Transactions on Geoscience and Remote Sensing, 57(1): 574–586.

Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A. C.; Lo, W.Y.; Dollar, P.; and Girshick, R. 2023. Segment Anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 4015–4026. Kuang, Z.; Chang, H.; Liang, B.; Wang, H.; He, L.; Li, F.; and Bi, H. 2026. MemOVCD: Training-Free Open-Vocabulary Change Detection via Cross-Temporal Memory Reasoning and Global-Local Adaptive Rectification. arXiv preprint arXiv:2604.26774. Lerebourg, M.; de Lange, F. P.; and Peelen, M. V. 2024. Attentional guidance through object associations in visual cortex. Science Advances, 10(41): eado6226. Li, D.; Wang, M.; Xiao, J.; and Yang, B. 2026a. Perspectives on spatio-temporal intelligence. ISPRS Journal of Photogrammetry and Remote Sensing, 232: 305–318. Li, K.; Cao, X.; Deng, Y.; Pang, C.; Xin, Z.; Qiao, H.; Gong, T.; Meng, D.; and Wang, Z. 2026b. DynamicEarth: How Far Are We from Open-Vocabulary Change Detection? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 6279–6287. Li, K.; Cao, X.; Deng, Y.; Song, J.; Liu, J.; Meng, D.; and Wang, Z. 2025a. SemiCD-VL: Visual-Language Model Guidance Makes Better Semi-Supervised Change Detector. IEEE Transactions on Geoscience and Remote Sensing, 63: 1–13. Li, K.; Liu, R.; Cao, X.; Bai, X.; Zhou, F.; Meng, D.; and Wang, Z. 2025b. SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10545–10556. Li, Z.; Tang, C.; Liu, X.; Li, C.; Li, X.; and Zhang, W. 2024. MS-Former: Memory-Supported Transformer for Weakly Supervised Change Detection With Patch-Level Annotations. IEEE Transactions on Geoscience and Remote Sensing, 62: 1–13. Liu, F.; Chen, D.; Guan, Z.; Zhou, X.; Zhu, J.; Ye, Q.; Fu, L.; and Zhou, J. 2024. RemoteCLIP: A Vision Language Foundation Model for Remote Sensing. IEEE Transactions on Geoscience and Remote Sensing, 62: 1–16. Liu, M.; Chai, Z.; Deng, H.; and Liu, R. 2022. A CNNTransformer Network With Multiscale Context Aggregation for Fine-Grained Cropland Change Detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15: 4297–4306. Oquab, M.; Darcet, T.; Moutakanni, T.; Vo, H. V.; Szafraniec, M.; Khalidov, V.; Fernandez, P.; Haziza, D.; Massa, F.; ElNouby, A.; Assran, M.; Ballas, N.; Galuba, W.; Howes, R.; Huang, P.-Y.; Li, S.-W.; Misra, I.; Rabbat, M.; Sharma, V.; Synnaeve, G.; Xu, H.; Jégou, H.; Mairal, J.; Labatut, P.; Joulin, A.; and Bojanowski, P. 2024. DINOv2: Learning Robust Visual Features without Supervision. Transactions on Machine Learning Research. Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable

Visual Models from Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139, 8748–8763. PMLR. Seydi, S. T. 2025. Assessment of the January 2025 Los Angeles County Wildfires: A Multi-Modal Analysis of Impact, Response, and Population Exposure. arXiv preprint arXiv:2501.17880. Siméoni, O.; Vo, H. V.; Seitzer, M.; Baldassarre, F.; Oquab, M.; Jose, C.; Khalidov, V.; Szafraniec, M.; Yi, S.; Ramamonjisoa, M.; Massa, F.; Haziza, D.; Wehrstedt, L.; Wang, J.; Darcet, T.; et al. 2025. DINOv3. arXiv preprint arXiv:2508.10104. Su, Y.; Song, Y.; Chen, J.; and Wen, Z. 2026. Seg2Change: Adapting Open-Vocabulary Semantic Segmentation Model for Remote Sensing Change Detection. arXiv preprint arXiv:2604.11231. Tan, X.; Chen, G.; Wang, T.; Wang, J.; and Zhang, X. 2024. Segment Change Model (SCM) for Unsupervised Change Detection in VHR Remote Sensing Images: A Case Study of Buildings. In IGARSS 2024 – 2024 IEEE International Geoscience and Remote Sensing Symposium, 8577–8580. Tan, X.; Chen, G.; Zhang, X.; Wang, T.; Wang, J.; Wang, K.; and Miao, T. 2025. TripleS: Mitigating Multi-Task Learning Conflicts for Semantic Change Detection in High-Resolution Remote Sensing Imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 230: 374–401. Tang, W.; Sun, H.; Li, Z.; Wang, Y.; and Zhang, F. 2026. CoRegOVCD: Consistency-Regularized Open-Vocabulary Change Detection. arXiv preprint arXiv:2604.02160. Wang, G.; Wu, C.; Zhang, X.; Peng, Z.; Lai, J.; Zhang, T.; and Tang, X. 2026. Seeking Consensus: GeometricSemantic On-the-Fly Recalibration for Open-Vocabulary Remote Sensing Semantic Segmentation. arXiv preprint arXiv:2604.26221.

Wang, Z.; Yi, J.; Chen, A.; Chen, L.; Lin, H.; and Xu, K. 2025. Accurate semantic segmentation of very high-resolution remote sensing images considering feature state sequences: From benchmark datasets to urban applications. ISPRS Journal of Photogrammetry and Remote Sensing, 220: 824–840. Yang, K.; Xia, G.-S.; Liu, Z.; Du, B.; Yang, W.; Pelillo, M.; and Zhang, L. 2022. Asymmetric Siamese Networks for Semantic Change Detection in Aerial Images. IEEE Transactions on Geoscience and Remote Sensing, 60: 1–18. Yang, Z.; Pei, G.; Yao, Y.; Zhou, T.; Ding, L.; and Shen, F. 2025. ChangeTitans: Toward Remote Sensing Change Detection With Neural Memory. IEEE Transactions on Geoscience and Remote Sensing, 63: 1–14. Zhang, C.; Yue, P.; Tapete, D.; Jiang, L.; Shangguan, B.; Huang, L.; and Liu, G. 2020. A Deeply Supervised Image Fusion Network for Change Detection in High Resolution Bi-Temporal Remote Sensing Images. ISPRS Journal of Photogrammetry and Remote Sensing, 166: 183–200. Zhang, L.; Zhang, L.; He, W.; Yuan, Q.; Zhao, L.; and Zhang, L. 2026a. From Deep Learning to Earth-Scale Intelligence: A Roadmap for Universal Large Remote Sensing Models. Proceedings of the IEEE, 1–23. Zhang, X.; Li, D.; Xia, Y.; Dong, X.; Yu, H.; Wang, J.; and Li, Q. 2026b. OmniOVCD: Streamlining OpenVocabulary Change Detection with SAM 3. arXiv preprint arXiv:2601.13895. Zheng, Z.; Zhong, Y.; Zhang, L.; and Ermon, S. 2024. Segment Any Change. In Advances in Neural Information Processing Systems (NeurIPS), volume 37, 12345–12367. Zhu, Y.; Li, L.; Chen, K.; Liu, C.; Zhou, F.; and Shi, Z. 2025. Semantic-CD: Remote sensing image semantic change detection towards open-vocabulary setting. In IGARSS 20252025 IEEE International Geoscience and Remote Sensing Symposium, 6388–6392. IEEE. Zhu, Z.; and Yang, B. 2025. UniVCD: A New Method for Unsupervised Change Detection in the Open-Vocabulary Era. arXiv preprint arXiv:2512.13089.

CogVis: Must Open-Vocabulary Change Detection Perceive the Scene Anew for Every Query? Supplementary Material

This supplement provides the implementation and evaluation details omitted from the main paper. Section A specifies the benchmark interfaces and prompt vocabularies. Section B expands the compact main-paper equations into an executable, fully parameterized description. Section C reports class-resolved results together with additional queryand region-level analyses, and Section D examines the same fixed inference rule in kilometer-scale scenes.

A

Experimental Protocol

Benchmarks and Evaluation Protocol

The seven benchmarks span semantic, binary, and damagestate change detection. SECOND (Yang et al. 2022) and SCSCD (Tan et al. 2025) require one semantic-change mask per foreground query; SCSCD is evaluated with its seven foreground classes. CLCD (Liu et al. 2022) merges the queried land-cover masks into a binary map. DSIFN (Zhang et al. 2020), LEVIR-CD (Chen and Shi 2020a), and WHU-CD (Ji, Wei, and Lu 2019) instead treat buildings as foreground. WHU-CD is the bitemporal change subset of the WHU Building Dataset. xBD (Gupta et al. 2019) evaluates building support and four post-event damage states. SECOND and SCSCD use macro-averaged metrics over foreground classes, binary benchmarks use foreground IoU and F1, and xBD uses five-state macro metrics that include background.

A.2

Evaluation Metrics

Let Yc and Ybc denote the reference and predicted pixel sets for class c. The corresponding confusion terms are TPc = |Yc ∩ Ybc |, FPc = |Ybc \ Yc |,

FNc = |Yc \ Ybc |.

(S2)

c∈Ceval

All experiments used 512 × 512 inputs and a dataset-specific prompt vocabulary that remained fixed across the complete test split. The adapters were trained and semantic memory was constructed using only CA-CDD; no target-benchmark test image contributed to optimization or memory construction. The resulting protocol evaluates cross-benchmark transfer under fixed text-specified vocabularies.

A.1

Class-wise and macro-averaged scores are TPc IoUc = , TPc + FPc + FNc 2TPc F1c = , 2TPc + FPc + FNc X 1 IoUc , mIoU = |Ceval | c∈Ceval X 1 mF1 = F1c . |Ceval |

(S1)

For xBD, predictions are evaluated in the standard fivestate output space comprising background, no damage, minor damage, major damage, and destroyed. The reported mIoU and mF1 are macro-averaged over these five states after final background assignment. To resolve performance by damage category, we also report the F1 score of each damage state on reference building pixels before background assignment. The same class-wise protocol is applied to every method in the comparison. Efficiency was measured after model initialization with a batch size of one. For N image pairs processed in time T , latency and throughput are T Latency = , N (S3) N FPS = . T Peak memory denotes the maximum GPU allocation during inference, and parameter counts are reported in millions. Query-scaling measurements used one warm-up and one synchronized full-request measurement for each vocabulary size, with model loading and image-disk input excluded. The profiling environment used Python 3.10.20, PyTorch 2.5.1 with CUDA 12.1 and an RTX 3090 GPU; inference was measured in FP32 with TF32 disabled. Evaluation scope. All results are evaluated over the complete test splits using fixed checkpoints and benchmarkstationary prompt vocabularies. Class-wise outcomes are interpreted jointly through absolute margins and crosscategory consistency. Multi-query computational accounting. CogVis separates computation shared by an image pair from computation repeated for each query. For a vocabulary of K queries, latency is T (K) = Tpair + K Tquery , (S4) where Tpair comprises frozen feature extraction, SCP, and pair embeddings, while Tquery comprises semantic decoding, memory retrieval, and region verification. Equation (S4)

Dataset

Query class Background Water

SECOND

Surface Low vegetation Tree Building Playground Bareland

SCSCD

Water Building Structure Farmland Vegetation Road

Text prompts {background} {water body, river, lake, reservoir} {bare land, bare soil, impervious barren ground, exposed ground} {low vegetation, grass, lawn, shrub, grassland, meadow} {tree canopy, forest, woodland, trees} {building, roof, house, structure} {playground, sports field, running track, athletic field} {bare land, bare soil, exposed ground, barren land} {water body, river, pond, canal} {building, rooftop, house, urban structure} {greenhouse, construction structure, built structure} {farmland, cropland, agricultural field} {vegetation, tree canopy, grassland, forest} {road, paved road, transportation road}

Table S1: Prompt vocabulary for semantic change detection. Braces group phrases whose responses are aggregated into one class map.

Dataset

A.3

Building

B

Model Specification

CogVis aligns each learned component with a distinct decision scale. The Scene Change Perceptron (SCP) extracts temporal evidence once per image pair. The Semantic Memory Calibrator (SMC) converts frozen semantic responses into a query-specific response and threshold. The Adaptive Region Filter (ARF) evaluates connected candidates at region level. We next define the full computation and the objectives for the Change, Score, and Gate Adapters.

Background

DSIFN

Building

LEVIR-CD

WHU-CD

Background Building Background Building

Text prompts {bare land, bare soil, exposed ground} {low vegetation, grassland} {road, paved road} {tree canopy, forest} {water body, river, lake} {cropland, farmland, agricultural field} {building, rooftop, urban structure} {ground, road, vegetation, farmland, bare ground, water} {urban building, building, building rooftop, roof, built-up area, urban structure} {ground, road, pavement, impervious surface, vegetation, water} {building, roof} {ground, road, vegetation, asphalt, pavement} {building, roof}

Table S2: Prompt vocabulary for binary change detection. CLCD merges seven class maps into a binary map; DSIFN, LEVIR-CD and WHU-CD use Building as the foreground query.

xBD

Text Prompt Vocabulary

Each prompt vocabulary contains the canonical class name, common remote-sensing synonyms, and variants that describe distinct visual forms. The frozen prompt decoder encodes each phrase independently, and phrase responses within a class are combined by a pixelwise maximum. The vocabulary remains unchanged for every image in a benchmark, with no image-specific prompt selection. In Tables S1– S3, braces group phrases assigned to one class, and bold labels denote evaluated foreground queries.

Grass Road Tree Water Cropland

CLCD

Dataset

records the accounting used for the main-paper efficiency study without repeating its measured latency, throughput, and memory results.

Query class Bareland

Query class Building support No damage

Text prompts {building, roof, house, building footprint, standing building} {undamaged building, intact building, standing intact building} Minor damage {slightly damaged building, minor roof damage, limited visible damage} Major damage {heavily damaged building, severely damaged building, partially collapsed building, major structural damage} {collapsed building, destroyed Destroyed building, completely collapsed building, building rubble, debris of a building}

Table S3: Prompt vocabulary for building localization and damage assessment on xBD.

B.1

End-to-End Computation

For registered images I 1 , I 2 ∈ RH×W ×3 , the frozen visual encoder (Carion et al. 2025) produces t {Ftl }L l=1 = E(I ),

S12 =

t ∈ {1, 2},

{F1l , F2l }L l=1 , B, es , ed

.

(S5)

Here, B is the SCP prior, and es and ed are scene and temporal CLIP embeddings (Radford et al. 2021). The resulting pair state, S12 , is computed once and reused for all queries.

For phrase p at time t, the frozen prompt decoder returns a semantic probability map stp and a presence score ϖpt . When activated, the instance branch also returns masks mtp,k and confidences otp,k . The phrase-level, class-level target, and strongest competing responses are   t t t t t qp = max ϖp sp , max op,k mp,k ,

define absolute and semantic-transition evidence: ρ  directional t Qc e t = Qt ⊙ , Q c c Qtc + Ctc + ϵ ′

e2 − Q e t ⊙ (1 − Q e 1 + λdir max Q e t ), Dc = λabs Q c c c c ′ t̸=t

Rc = clip[0,1] (Dc ⊙ Gc ⊙ Sc pc ) . (S11)

k

Qtc = max qtp , p∈Pc

(S6)

Ctc = max max qtp . n∈Nc p∈Pn

t

The corresponding scalar class-presence score is πct = maxp∈Pc ϖpt . SMC maps the endpoint responses and B to a continuous change response Rc and threshold τc . ARF filters f c = 1[Rc ≥ τc ]: the connected components of M SCP

SMC

ARF

(I 1 , I 2 ) −−−→ B −−−→ (Rc , τc ) −−−→ Mc . c

B.2

(S7)

Change Adapter and Scene Change Perceptron

Construction. Here Catch (·) denotes channel-wise concatenation along the feature dimension, and Hchg (X, ∆) denotes the change head, a shallow convolutional network (see Table S4) that maps the paired descriptor and temporal residual to a spatial change logit map. SCP operates on the second level of the frozen feature pyramid, Ft = Ft2 , and constructs a paired descriptor from endpoint appearance, absolute discrepancy and feature agreement:
X = Catch F1 , F2 , |F2 − F1 |, F1 ⊙ F2 , (S8) ∆ = W↑ ϕ(W↓ X). The low-rank Change Adapter produces the temporal residual ∆. A temporal gate limits its root-mean-square magnitude to 0.25 of the frozen feature magnitude. The change head then predicts

B = σ Up Hchg (X, ∆) . (S9) The category-agnostic prior is computed once per image pair and reused for all queries. The residual contributes only to B; semantic decoding uses the unmodified frozen feature pyramid. Optimization. The Change Adapter is trained on binary source masks with equally weighted class-balanced binary cross-entropy and soft Dice losses: Lchg = Lbal-bce (B, Y) + Ldice (B, Y).

(S10)

The BCE positive-class weight is the mini-batch negativeto-positive pixel ratio clipped to [1, 20]. Source masks are resized to the change-logit resolution by nearest-neighbor interpolation.

B.3

The response modulation terms are Gc = ϵg + (1 − ϵg )B,   e t ) ⊙ (1 − B) , Sc = clip[0,1] 1 − λs min(Q

Semantic Memory Calibrator and Score Adapter

Semantic-change response. Prompt competition downweights target activation that is also explained by nearby semantic classes. The competition-adjusted endpoint responses

c

(S12)

pc = |πc2 − πc1 |. The fixed coefficients are ϵ = 10−6 , ρ = 2, λabs = 0.55, λdir = 0.25, ϵg = 0.10, and λs = 0.45. Thus, Rc retains query-specific transitions only when semantic transition, shared temporal support, stable-region suppression, and endpoint presence change agree. Memory retrieval. All spatial maps are bilinearly resized to the 512 × 512 inference grid. For a map X, let Qq (X) denote its spatial q-quantile and let Mass>a (X) be the fraction of pixels whose value exceeds a. The 20-dimensional response statistic is rc = [µ, σ, Q0.50 , Q0.75 , Q0.85 , Q0.90 , Q0.95 , Q0.98 , max, Mass>µ , Mass>Q0.85 ](Rc ), b = [µ, σ, Q0.75 , Q0.85 , Q0.95 , max](B), uc = [µ, Q0.85 ](Rc ⊙ B), ξ c = [rc , b, uc , πcmax ] ∈ R20 , (S13) where πcmax = max(πc1 , πc2 ). CLIP ViT-B/16 provides normalized endpoint embeddings v1 , v2 ∈ R512 , from which   v1 + v2 es = Norm , 2 (S14) ed = Norm(|v2 − v1 |). The query embedding ec is the normalized embedding of the first registered phrase for class c. With Norm(x) = x/(∥x∥2 + ϵ), the retrieval key is kc = Norm([es , 0.70ed , ec , 0.25ξc ]) ∈ R1556 . (S15) The memory enumerates 4,968 CA-CDD pairs and 12 registered semantic anchors, producing 59,616 rows. The source operating point is selected over Gτ = {0.02, 0.04, . . . , 0.80} (S16) by τi⋆ = arg max IoU(1[Ri ≥ τ ], Yi ) . (S17) τ ∈Gτ

An empty source mask is assigned τi⋆ = 0.02, and exact ties are resolved by the smallest threshold. Cosine similarity retrieves the top 16 source-memory entries, denoted Kc . After subtracting the largest similarity for numerical stability, temperature weighting with Tm = 0.07 gives exp((si − smax )/Tm ) αic = P , j∈Kc exp((sj − smax )/Tm ) (S18) X τ̄c = αic τi⋆ , i∈Kc

where si = k⊤ c ki . Confidence-aware threshold correction. The label-free fallback Tadp is deterministic. Let O(R) denote 128-bin global Otsu thresholding. An edge-aware value Oedge (R) averages this threshold with Otsu computed on a one-step 3 × 3 dilation of the top-20% Sobel-gradient support. It reverts to O(R) when the maximum gradient is below 10−8 or the dilated support contains fewer than 64 pixels. Then Tadp (R) = clip[0,1] median{Oedge (R),
Q0.85 (R), µ(R) + σ(R)} .

(S19)

The Score Adapter receives the 20 response statistics, memory anchor, top-1 similarity, and mean retrieved similarity as a 23-dimensional input: zc = [ξ c , τ̄c , smax , smean ], c c ∆τc = 0.20 tanh Ascore (zc ),   1 X (−i) Lscore = ℓsl1 ∆τi , τi⋆ − τ̄i , N i τc = clip[0.02,0.80] (max {τ̄c + ∆τc , 0.75Tadp (Rc )}) . (S20) Leave-one-out retrieval removes only the current row and retains all remaining rows, including the other semantic anchors from the same source pair. At inference, the memory anchor transfers a source operating point, the bounded correction adapts it to the current image–query pair, and the deterministic scaled adaptive estimate 0.75Tadp (Rc ) supplies a conservative lower bound.

B.4

Adaptive Region Filter and Gate Adapter

Region representation. Thresholding Rc yields Ωc = f c ). For component ωj , let aj = |ωj |, bj = CCL(M | bbox(ωj )|, and A = HW . Here, CCL denotes 8-connected component labeling, bbox the tight bounding box, and Top0.20,j the mean of the largest max(1, ⌊0.20aj ⌋) component values. Statistics restricted to ωj are summarized by rj (X) = [µj (X), σj (X), max(X), Top0.20,j (X)], j

Massj (X; θ) = a−1 j |{x ∈ ωj : X(x) > θ}|. (S21) With θB = Tadp (B), the four descriptor blocks are q φgeo = [a /A, aj /A, bj /A, aj /bj ], j j φsem = rj (Rc ), j φtmp = [rj (B), Massj (B; θB )], j φcal j = [Massj (Rc ; τc ), µj (Rc ⊙ B), τc , θB , Top0.20,j (Rc ) − τc , Top0.20,j (B) − θB ]. (S22) Their concatenation gives φj = tmp sem cal 19 Cat(φgeo , φ . The four blocks , φ , φ ) ∈ R j j j j encode geometry, semantic confidence, temporal support, and calibrated margin. Gate Adapter optimization. The Gate Adapter predicts uj = σ(Agate (φj )). Source components are labeled from

precision- and recall-oriented overlap with the binary source mask: |ωj ∩ Y| qjp = , |ωj | |ωj ∩ Y| qjr = , |Y| + ϵ h yj = 1 |ωj ∩ Y| ≥ nmin (S23) i ∧ (qjp ≥ δp ∨ qjr ≥ δr ) , 1 Xh w+ yj log uj Lgate = − NΩ j i + (1 − yj ) log(1 − uj ) . Training uses nmin = 8, δp = 0.30, δr = 0.05, and w+ = 2.0454. Components outside [8, 262,144] pixels are excluded from Gate training. These constants are fixed before target-benchmark evaluation. During inference, large coherent regions are retained. All other components must satisfy the learned reliability criterion:  [  |ωj | Mc = ωj ≥ ηa ∨ uj ≥ τr . (S24) HW ωj ∈Ωc

Filtering retains or removes complete components and leaves pixels within each retained component unchanged.

B.5

Architecture, Optimization, and Preprocessing

Table S4 specifies every trainable operator. The visual encoder, prompt decoder, and CLIP encoders remain frozen. The three adapters are stored as separate checkpoints. All optimizers use β1 = 0.9, β2 = 0.999, and ϵopt = 10−8 . Change Adapter training clips the global gradient norm at 1.0. Images are resized bilinearly to 512 × 512, and masks use nearest-neighbor resampling. Paired feature jitter uses replicate padding, a crop displacement of at most four pixels at the highest loaded FPN resolution, scaleconsistent offsets across levels, and horizontal flipping with probability 0.5. Both timestamps receive the same transform. The Change Adapter is optimized first; its frozen response archive is then used to build the memory. The Score and Gate Adapters are fitted on the resulting queryand component-level representations. All fixed inference values are shared across target benchmarks: Kmem = 16, Tm = 0.07, [τmin , τmax ] = [0.02, 0.80], correction bound 0.20, adaptive floor coefficient 0.75, τr = 0.60, ηa = 0.04, and an SCP RMS cap of 0.25.

C

Extended Results

This section adds evidence not contained in the main paper. Class-resolved analyses identify where the aggregate gains occur on SECOND, SCSCD, and xBD. Additional experiments then quantify query-specific threshold selection and region-level cue usage, while the qualitative plates expand the number of examples shown for all seven benchmarks.

Module

Input

Ordered operators

SCP adapter

1024 × H2 × W2

SCP head and gain

Four spatial maps

Score Adapter

zc ∈ R23

Gate Adapter

φj ∈ R19

Conv1×1 1024 → 16, GELU, Conv1×1 16 → 256; no bias; rank and scale 16 Conv3×3 4 → 96, GN8 , GELU, Conv3×3 96 → 96, GN8 , GELU, Conv1×1 96 → 1, scalar gain Linear(23, 64), LN, ReLU, Dropout(0.05), Linear(64, 64), ReLU, Linear(64, 1) Linear(19, 64), LN, ReLU, Dropout(0.05), Linear(64, 64), ReLU, Linear(64, 1), sigmoid

Trainable parameters 20,480 86,882 5,889 5,633

Table S4: Trainable architecture of SCP, SMC, and ARF. The counts cover the executed adapters and heads rather than the frozen foundation models. Comparison methods. We included UCD-SCM (Tan et al. 2024), AnyChange (Zheng et al. 2024), DynamicEarth (Li et al. 2026b), UniVCD (Zhu and Yang 2025), OmniOVCD (Zhang et al. 2026b), AdaptOVCD (Dou et al. 2026), OpenDPR (Guo et al. 2026), Seg2Change (Su et al. 2026), and CoRegOVCD (Tang et al. 2026). Inst-CEG is the instance-level CEG baseline from SemiCD-VL (Li et al. 2025a). SAM3-CD denotes the SAM 3 baseline (Carion et al. 2025) implemented in the UniVCD study (Zhu and Yang 2025). Tables S5–S7 separate benchmark outcomes from mechanism evidence. Baseline results follow the benchmarkaligned comparison used in the main paper, and CogVis results are computed over the complete target test splits using the definitions in Section A. Table S8 adds query- and regionlevel measurements that are not tabulated in the main paper. For each example, every method is evaluated on the same image pair.

C.1

Detailed Results on Multi-Class Benchmarks

SECOND evaluates direction-invariant class involvement,   YcSECOND (x) = 1 Y 1 (x) = c ∨ Y 2 (x) = c , (S25) and scores the six foreground masks independently before macro averaging. SCSCD evaluates seven foreground change categories, while xBD evaluates the complete five-state damage map. Together, these protocols test whether the overall gains extend across semantic classes and output spaces. SECOND. Table S5 resolves the main-paper aggregate by foreground category. CogVis ranks first for Building, Tree, Low vegetation, and Surface and is within 0.02 points of the best Water IoU. These gains span objects, vegetation, and background-like surfaces rather than one dominant class. Playground remains difficult because its mixed impervious and vegetated appearance increases semantic overlap and fragments connected support. SCSCD. Table S6 shows that CogVis leads on Water and Vegetation and ranks second on Bareland, Building, Farmland, and Road. Farmland and Structure exhibit stronger cross-class ambiguity because agricultural structures, greenhouse-like regions, and built surfaces share closely related visual patterns.

xBD. Table S7 complements the main-paper mIoU with state-wise F1 and full-map mF1. CogVis obtains the strongest F1 for no damage, major damage, and destroyed buildings. The advantage is most pronounced for major and destroyed structures, where temporal support and region verification are particularly important. Minor damage remains difficult because its visual evidence lies between the adjacent nodamage and major-damage states.

C.2

Additional Calibration and Region Analyses

To avoid duplicating the analyses already reported in the main paper, this supplement adds only two complementary measurements. Panel (a) isolates query-level threshold selection, and panel (b) measures how each ARF evidence family contributes to candidate ranking. Query-specific threshold calibration. We compare a fixed threshold, the deterministic label-free rule, memory transfer, and the Score Adapter on SECOND and CLCD. The Score Adapter achieves the highest SECOND mIoU and CLCD IoU (Table S8, panel (a)). Its two-dataset mean exceeds the fixed threshold by 5.94 points and the memory anchor by 1.81 points, indicating that the learned correction complements rather than merely reproduces the retrieved operating point. ARF evidence ablation. Removing any evidence family reduces the area under the precision–recall curve (AUPRC) and the area under the ROC curve (AUROC) on SECOND and CLCD (Table S8, panel (b)). Temporal support produces the largest mean decrease, followed by geometry and semantic confidence. These ranking results indicate that the Gate Adapter uses complementary cues rather than relying on a single component statistic.

C.3

Qualitative Comparison

The main paper provides a consolidated qualitative overview. Figures S1 and S2 expand that evidence with four pairs from each of SECOND, SCSCD, and xBD and three pairs from each of WHU-CD, LEVIR-CD, CLCD, and DSIFN. For each example, every method is evaluated on the same image pair; the additional samples expose category-specific and nuisance-driven failure patterns rather than repeat the main-paper examples.

Method

Class-wise IoU / F1

Mean

Building Tree Water Low veg. Surface Play. mIoU / mF1 UCD-SCM 22.06/36.15 3.36/6.50 1.11/2.20 10.35/18.77 12.88/22.83 0.66/1.31 8.40/14.63 AnyChange 27.22/42.79 3.89/7.49 1.11/2.19 14.17/24.82 24.15/38.90 0.52/1.03 11.84/19.54 Inst-CEG 23.66/38.27 9.64/17.59 10.03/18.24 0.57/1.14 0.00/0.00 18.84/31.70 10.46/17.82 DynamicEarth-I-M-C 31.74/48.19 10.89/19.64 12.34/21.97 0.40/0.79 0.00/0.00 26.93/42.43 13.72/22.17 DynamicEarth-M-C-I 38.60/55.70 15.54/26.89 15.37/26.64 21.25/35.06 28.04/43.80 22.67/36.97 23.58/37.51 SAM3-CD 22.60/36.90 12.40/22.10 5.00/9.60 18.20/30.90 26.70/42.10 11.80/21.10 16.12/27.12 UniVCD 43.20/60.40 18.90/31.90 8.20/15.20 24.90/39.80 28.00/43.70 0.00/0.00 20.53/31.83 OmniOVCD 45.20/62.30 16.70/28.60 21.20/35.00 24.50/39.30 27.70/43.40 27.00/42.40 27.05/41.83 AdaptOVCD 46.85/63.81 12.67/22.49 23.34/37.84 22.15/36.27 33.99/50.74 29.28/45.30 28.05/42.74 OpenDPR 42.40/59.50 20.90/34.50 17.50/29.80 23.00/37.40 30.20/46.40 37.30/54.30 28.55/43.65 Seg2Change 61.76/76.36 16.10/27.73 22.88/37.23 18.69/31.49 37.26/54.29 17.81/30.23 29.08/42.89 CoRegOVCD 48.91/65.69 20.93/34.61 29.88/46.01 27.71/43.40 32.27/48.79 30.31/46.52 31.67/47.50 CogVis (Ours) 63.07/77.35 24.55/39.43 29.86/45.99 30.89/47.20 39.76/56.90 6.73/12.61 32.48/46.58

Table S5: Class-wise comparison on SECOND. Each entry reports IoU/F1; Mean is the unweighted average across six foreground classes. All values are percentages. Bold and underlined entries indicate the best and second-best results, respectively. Method

Class-wise IoU / F1

Mean

Bare Water Building Struct. Farm Veg. Road mIoU / mF1 UCD-SCM 8.77/16.13 1.69/3.32 8.11/15.00 10.09/18.33 4.81/9.18 9.01/16.52 3.06/5.95 6.51/12.06 9.32/16.51 AnyChange 16.29/28.01 2.16/4.22 11.03/19.88 10.07/18.30 5.93/11.19 17.85/30.29 1.89/3.71 3.68/6.90 Inst-CEG 0.00/0.00 7.92/14.68 5.45/10.34 0.00/0.00 0.00/0.00 6.79/12.71 5.58/10.58 DynamicEarth-I-M-C 0.00/0.00 12.34/21.97 13.55/23.87 0.00/0.00 0.00/0.00 0.00/0.00 10.22/18.54 5.16/9.19 DynamicEarth-M-C-I 22.59/36.85 33.14/49.78 28.55/44.42 13.21/23.33 5.48/10.39 19.76/33.01 3.08/5.97 17.97/29.11 Seg2Change 27.51/43.15 21.89/35.92 47.20/64.13 0.77/1.53 11.40/20.47 29.84/45.97 23.91/38.60 23.22/35.68 CogVis (Ours) 27.28/42.86 40.92/58.08 34.33/51.12 9.49/17.33 9.35/17.10 35.84/52.76 20.33/33.80 25.36/39.01

Table S6: Class-wise comparison on SCSCD. Each entry reports IoU/F1; Mean is the unweighted average across seven foreground classes. All values are percentages. Bold and underlined entries indicate the best and second-best results, respectively. Method Seg2Change OmniOVCD UCD-SCM OpenDPR AnyChange AdaptOVCD DynamicEarth-M-C-I CogVis (Ours)

Damage-State F1

Full Map

No Damage Minor Major Destroyed mIoU mF1 86.90 0.64 0.79 2.34 17.94 20.04 18.92 21.37 87.06 1.01 0.02 9.50 86.94 3.07 1.86 8.75 19.04 21.62 87.31 3.26 3.67 12.57 19.58 22.63 87.18 4.74 3.30 13.01 19.63 22.70 87.46 0.50 0.00 39.30 22.07 26.22 87.41 0.01 0.00 35.42 24.51 29.30 88.06 0.86 11.69 50.83 32.68 40.63

Table S7: Class-wise comparison on xBD. The four damage-state columns report F1 on reference building pixels before final background assignment. Full-map mIoU and mF1 are macro-averaged over background and the four damage states after background assignment. All values are percentages; bold and underlined entries indicate the best and second-best results, respectively. SECOND and SCSCD are rendered with a symmetric class-involvement target Mcvis = MCD ∧ [(Y 1 = c) ∨ (Y 2 = c)].

(S26)

UCD-SCM and AnyChange produce category-agnostic masks. A common frozen open-vocabulary classifier therefore assigns semantic labels within their predicted support. Native semantic outputs are rendered directly. Across the semantic examples, CogVis confines Building,

Water, Vegetation, and Surface responses to object-aligned regions while preserving their class identity. In the binary panel, its clearest visible advantage is the rejection of broad false-positive masks that saturate stable roofs, fields, or background in several category-agnostic outputs. The retained responses instead follow changed structures in the post-event image. On xBD, building support and severe-damage regions are delineated more coherently, consistent with the full-map

(a) Query-specific threshold calibration Threshold rule Fixed global Label-free adaptive Memory anchor Score Adapter

SECOND mIoU CLCD IoU Mean 23.30 15.19 19.25 23.06 17.97 20.52 28.81 17.94 23.38 31.12 19.26 25.19

(b) ARF evidence ablation Metric SECOND ∆AUPRC SECOND ∆AUROC CLCD ∆AUPRC CLCD ∆AUROC Mean drop

Semantic Temporal Geometry −0.0275 −0.0105 −0.0237 −0.0055 −0.0340 −0.0055 −0.0147 −0.0742 −0.0500 −0.0056 −0.0887 −0.0033 0.0133 0.0519 0.0206

Table S8: Supplementary mechanism evidence not tabulated in the main paper. (a) Query-level threshold calibration; Mean averages SECOND mIoU and CLCD IoU. (b) Region-level ARF evidence ablation under the source-trained overlap criterion. In panel (a), bold and underlined values denote the best and second-best results, respectively; boldface in panel (b) highlights the largest mean drop. gains in Table S7. These displayed patterns are consistent with the combined effects of pair-level suppression, queryspecific calibration, and component-level filtering.

C.4

Category-Level Interpretation

The remaining errors concentrate in categories with mixed visual composition or fine-grained semantic transitions. Playground combines impervious surfaces with vegetation, while Structure overlaps visually with buildings and agricultural structures. Damage grades with subtle appearance differences require finer separation than the endpoint states. These patterns indicate that the remaining performance gap is primarily associated with fine-grained semantic discrimination. In contrast, the consistent gains on Building, Water, Vegetation, Surface, and the complete xBD map demonstrate that the scope-aligned design remains effective across diverse spatial structures and change types.

D

Large-Scene Wildfire Damage Mapping

Application setting. We mapped registered pre- and postevent RGB imagery from the January 2025 Palisades and Eaton fires in Los Angeles County (Seydi 2025) using the trained CogVis model and xBD prompt vocabulary without site-specific fine-tuning. The aligned Palisades pair contains 17,592×13,442 pixels at approximately 0.51 m ground sampling distance and covers 8.98 × 6.86 km. The Eaton pair contains 12,721 × 10,455 pixels at approximately 0.55 m ground sampling distance and covers 7.01 × 5.76 km. Both pairs use EPSG:32611 and a shared within-scene affine transform. Because site-specific reference labels are unavailable, this case study supports only qualitative spatial transfer, not local accuracy or damage-extent estimates. Mapping procedure. We partitioned each scene into 512 × 512 tiles at a 384-pixel stride, yielding 1,610 Palisades tiles and 891 Eaton tiles. For tile prediction ybt (x) and class c, overlap was fused by P wt (x) 1[b yt (x) = c] Vc (x) = t∋x P , (S27) t∋x wt (x) yb(x) = arg max Vc (x), c

where wt is a separable two-dimensional Hann window with a 10−3 border floor. The weighted vote suppresses tile-edge discontinuities while retaining a defined estimate at the scene boundary.

Geospatial sampling and rendering. Figure S3 overlays the fused maps on post-event RGB imagery. Three 1024 × 1024-pixel regions were sampled at fixed relative positions in each scene and grid-adjusted to retain at least 99.9% paired coverage. Panels (a)–(c) show Palisades, and panels (d)–(f) show Eaton. Overview maps use opaque class colors; detail panels use 0.72 opacity to retain roof, parcel, and street context. Prompt formulation. The standalone phrases in Table S3 were applied verbatim to both scenes. Five building-support phrases localized candidate structures; four state-specific banks distinguished no damage, minor damage, major damage, and destroyed. Phrases were encoded independently, and responses within each state were aggregated by pixelwise maximum, preserving the vocabulary and inference rule evaluated on xBD. Spatial patterns. Palisades contains broad, contiguous destroyed-state responses. Panels (a) and (b) resolve these responses along roof rows and residential blocks; panel (c) instead shows a predominantly no-damage area interspersed with smaller destroyed regions. Eaton is dominated by nodamage responses. Panels (d) and (e) isolate smaller destroyed clusters within the urban fabric, whereas panel (f) contains the most compact Eaton response. Across 2,501 tiles, the fused predictions preserve roof- and block-level spatial continuity without obvious seams at the displayed scale. These maps show how the frozen model and prompt vocabulary can be applied to large geospatial scenes through overlap-aware tiled inference. Figure S3 provides a scenescale visualization of the same fixed inference rule evaluated on xBD.

T2

GT

UCD-SCM

AnyChange

DynamicEarth-MCI AdaptOVCD

OpenDPR

OmniOVCD

Seg2Change

CogVis (Ours)

xBD

SECOND

SCSCD

T1

Change / Building Water Vegetation / Tree

Farmland / Grass Bareland / Ground Structure

Road Playground No-Damage

Minor-Damage Major-Damage Destroyed

Figure S1: Qualitative comparison on SCSCD, SECOND, and xBD. Four examples are shown per dataset. Columns show T1 , T2 , ground truth, seven baselines, and CogVis; colors denote semantic change classes on SCSCD and SECOND and damage states on xBD.

T2

GT

UCD-SCM

AnyChange

DynamicEarth-MCI AdaptOVCD

OpenDPR

OmniOVCD

Seg2Change

CogVis (Ours)

DSIFN

CLCD

LEVIR-CD

WHU-CD

T1

Changed region

Figure S2: Qualitative comparison on WHU-CD, LEVIR-CD, CLCD, and DSIFN. Three examples are shown per dataset. Columns show T1 , T2 , ground truth, seven baselines, and CogVis; orange denotes changed pixels overlaid on T2 .

Figure S3: Large-scene wildfire damage predictions for the Palisades and Eaton fire areas in California. Post-event overview maps show Palisades (left) and Eaton (right), with 1024 × 1024-pixel regions (a)–(c) and (d)–(f) enlarged below. Detail panels overlay predictions on RGB imagery. Olive green, tan, rose, and dark red denote no damage, minor damage, major damage, and destroyed, respectively; axis ticks report WGS 84 coordinates, and arrows indicate north.

Related documents

• GradeTrap: Authority Cues in Images Shift VLM Judgments Despite Explicit Instructions to Ignore Them#973668
• Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models#1117551
• Individual gaze predicts individual scene descriptions#411939
• The LAIA Dataset: Labelled Attention for Intelligent Automobiles#411137
• Elastic Threshold Attention: Learned Contextual Sparsity for Long-Context Decoding#1034862