Open-Vocabulary Change Detection (OVCD)

• Open-Vocabulary Change Detection is a remote-sensing task that uses arbitrary text prompts to localize and semantically identify changes between two temporal images.
• Key methodologies include decomposing the task into components like identification, comparison, and mask proposal, leveraging training-free and unsupervised pipelines.
• Recent approaches combine foundation models such as CLIP, SAM, and DINO to address dataset scarcity and improve robustness against variable imaging conditions.

Open-Vocabulary Change Detection (OVCD) is a remote-sensing change analysis task that localizes and semantically identifies changes between bi-temporal images under arbitrary user-defined text prompts rather than a fixed closed label set. In contrast to binary change detection, which asks only whether change occurred, and conventional semantic change detection, which assumes predefined categories, OVCD asks both where a queried concept changed and what semantic change occurred for concepts that may be unseen during task-specific training. The topic emerged from the convergence of change detection with vision-language and segmentation foundation models, especially CLIP-, SAM-, and DINO-family systems, and has rapidly diversified into training-free, unsupervised, weakly supervised, and adapter-based formulations for remote sensing (Li et al., 22 Jan 2025, Zhu et al., 12 Jan 2025).

1. Definition and emergence

DynamicEarth introduced OVCD as a task that bridges vision and language to detect changes across any category, formalizing the setting as change localization and identification between two temporally separated images $x_{img_1}$ and $x_{img_2}$ with arbitrary text labels $x_{text}$ (Li et al., 22 Jan 2025). Subsequent work sharpened the formulation around remote sensing. AdaptOVCD defined a zero-shot inference mapping

$$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$

where $I_a$ and $I_b$ are pre- and post-change images, $\mathcal{T}$ is the text prompt describing the target change category, and $M$ is the binary change mask (Dou et al., 6 Feb 2026). CoRegOVCD specified the queried-concept version with aligned images

$I_a, I_b \in \mathbb{R}^{H \times W \times 3}$

and output

$\hat{Y}^{(c)} \in \{0,1\}^{H \times W},$

indicating where the queried concept changes between the two dates (Tang et al., 2 Apr 2026).

The field developed out of dissatisfaction with closed-set change detection. Across the literature, conventional methods are described as dataset-dependent, annotation-hungry, and restricted to predefined categories, with poor generalization to unseen object types, diverse scenes, sensor geometries, and open-world user queries (Zhu et al., 15 Dec 2025, Dou et al., 6 Feb 2026). Semantic-CD represents an early transition point: it explicitly imported CLIP semantics into semantic change detection, but also noted that current SCD datasets are too small in category diversity for a fully standard OVCD benchmark, leading it to “degrade” the setting into an open-capability SCD model on SECOND (Zhu et al., 12 Jan 2025).

A recurring definitional distinction is that OVCD is not merely open-vocabulary segmentation applied independently at two timestamps. The problem is explicitly bi-temporal and must suppress pseudo-changes caused by illumination, season, atmospheric conditions, view shifts, radiometric inconsistency, and semantic ambiguity (Tang et al., 2 Apr 2026, Zhu et al., 18 Jun 2026). This makes OVCD a joint problem of semantic querying, temporal reasoning, and robust change validation rather than a direct extension of single-image open-vocabulary recognition.

2. Canonical decompositions of the task

A foundational contribution of the OVCD literature is the decomposition of the task into reusable subproblems. DynamicEarth argued that a complete OVCD system should include three components—Identifier, Comparator, and Mask Proposal—and instantiated two training-free paradigms: M-C-I (Mask Proposal → Comparator → Identifier) and I-M-C (Identifier → Mask Proposal → Comparator) (Li et al., 22 Jan 2025). The first discovers all potential changes and then classifies them; the second identifies targets of interest first and then determines whether their states have changed.

OpenDPR further formalized OVCD as a two-stage pipeline: change localization and category identification (Guo et al., 29 Mar 2026). Its controlled error attribution on Hi-UCD mini reported 49.1% mIoU for Oracle Identification, 23.8% mIoU for Oracle Change Proposal, and 13.1% mIoU for the full M-C-I pipeline, leading to the diagnosis that category identification errors dominate final performance degradation. This diagnosis shifted attention away from viewing OVCD as a monolithic problem and toward separate treatment of localization and recognition.

Other papers retained the same broad decomposition but instantiated it differently. AdaptOVCD wrote the pipeline as

$x_{img_2}$0

with $x_{img_2}$1 for instance segmentation, $x_{img_2}$2 for feature comparison, and $x_{img_2}$3 for semantic identification (Dou et al., 6 Feb 2026). ReA-OVCD, by contrast, treated dense pixel-wise semantic parsing as the source of candidate change regions and then introduced explicit semantic and spatial reliability checks to validate those candidates (Zhu et al., 18 Jun 2026). MemOVCD criticized “late-comparison systems,” writing them as

$x_{img_2}$4

and reformulated OVCD as conditional cross-temporal interpretation,

$x_{img_2}$5

thereby making temporal coupling itself part of the task definition (Kuang et al., 29 Apr 2026).

These decompositions are significant because they expose different failure modes. Proposal-first pipelines emphasize candidate coverage and region comparison; identifier-first pipelines emphasize category discovery; dense posterior methods emphasize calibrated semantic discrepancy; and memory-based methods emphasize temporally conditioned semantic continuity. A plausible implication is that OVCD has evolved less as a single algorithmic lineage than as a family of competing answers to the question of where semantic change evidence should first be made reliable.

3. Methodological families and representative systems

The OVCD literature now contains several distinct methodological families.

Method	Core design	Supervision mode
Semantic-CD	CLIP-based open semantic prompter with decoupled BCD and SCD decoders	Two-stage trained
DynamicEarth	Training-free M-C-I and I-M-C frameworks from off-the-shelf foundation models	Training-free
UniVCD	Frozen SAM2 + frozen CLIP + SCFAM + post-processing	Unsupervised
AdaptOVCD	ARA + ACT + ACF over SAM-HQ, DINOv3, DGTRS-CLIP	Training-free
OpenDPR / OpenDPR-W	Diffusion-guided prototype retrieval, optionally with S2C	Training-free / weakly supervised
OmniOVCD	Standalone SAM 3 with SFID	Training-free
CoRegOVCD	CPC + SPD + GeoGate + RCD dense posterior reasoning	Training-free
Seg2Change	CACH adapter plus OVSS model	Trained adapter
MemOVCD	Cross-temporal memory reasoning and global-local adaptive rectification	Training-free
ReA-OVCD	SCR + BCR over frozen SAM-3 parsing	Training-free

One family centers on CLIP-derived semantics. Semantic-CD used a bi-temporal CLIP visual encoder, an open semantic prompter, a binary change detection decoder, and a semantic change detection decoder, with fully decoupled multi-task learning to separate binary localization from semantic labeling (Zhu et al., 12 Jan 2025). UniVCD extended this line into unsupervised OVCD by combining a frozen SAM2 encoder, a frozen CLIP encoder, a lightweight SAM-CLIP Feature Alignment Module (SCFAM), and a post-processing stage, explicitly avoiding end-to-end fine-tuning of SAM2 or CLIP (Zhu et al., 15 Dec 2025).

A second family uses multi-model training-free pipelines. DynamicEarth assembled off-the-shelf foundations into M-C-I and I-M-C systems (Li et al., 22 Jan 2025). AdaptOVCD organized the same general idea into data, feature, and decision levels, with Adaptive Radiometric Alignment (ARA), Adaptive Change Thresholding (ACT), and Adaptive Confidence Filtering (ACF) layered over SAM-HQ, DINOv3, and DGTRS-CLIP (Dou et al., 6 Feb 2026). OpenDPR departed from text matching by building visual prototypes with GPT-4, Diffusion-Sat, APE, and DINOv2, then retrieving category labels in visual feature space; its optional OpenDPR-W added S2C, a weakly supervised spatial-to-change module trained with image-level binary change labels (Guo et al., 29 Mar 2026).

A third family pursues single-backbone or tighter-integrated designs. OmniOVCD argued that chaining CLIP, DINO, and SAM causes feature mismatch and instability, and instead used SAM 3 alone with Synergistic Fusion to Instance Decoupling (SFID) (Zhang et al., 20 Jan 2026). CoRegOVCD likewise favored dense inference over explicit multi-model association, converting raw concept responses into competition-aware posteriors and then regularizing them geometrically and regionally (Tang et al., 2 Apr 2026). Seg2Change adapted an open-vocabulary semantic segmentation model to change detection by coupling a trained category-agnostic change head (CACH) with an OVSS backbone, specifically SegEarth-OV3 in the main experiments (Su et al., 13 Apr 2026).

A fourth family emphasizes temporal coupling and reliability correction. MemOVCD reformulated bi-temporal change detection as a two-frame tracking problem using SAM 3.1, weighted bidirectional propagation, histogram-aligned transition frames, and global-local adaptive rectification (Kuang et al., 29 Apr 2026). ReA-OVCD began from pixel-wise semantic discrepancies produced by a frozen SAM-3 parser and then validated them with Semantic Change Reasoning (SCR) and Boundary-aware Change Refinement (BCR) (Zhu et al., 18 Jun 2026).

4. Core algorithmic mechanisms

Despite architectural diversity, several recurring mechanisms organize the technical core of OVCD.

Proposal-based comparison is prominent in early training-free systems. DynamicEarth’s ovcdred computes a mask-level change score with latent feature matching,

$x_{img_2}$6

after generating class-agnostic masks and applying NMS (Li et al., 22 Jan 2025). OpenDPR retains the proposal-then-identify logic but replaces text matching with prototype retrieval in visual space. For a proposal $x_{img_2}$7, it evaluates

$x_{img_2}$8

and reports that global-max usually performs better than category-mean because it preserves strong semantic matches and avoids averaging away discriminative sub-attributes (Guo et al., 29 Mar 2026).

Cross-modal semantic alignment defines another major line. UniVCD’s SCFAM aligns SAM2’s spatial detail with CLIP’s semantic priors through a layer-wise fusion block, lightweight adapters, and multiple projection heads. Change is then computed from per-image class-probability maps. For category $x_{img_2}$9,

$x_{text}$0

making change estimation category-wise rather than purely binary (Zhu et al., 15 Dec 2025). This category-wise formulation is what the paper identifies as the mechanism enabling a unified treatment of binary, semantic, and multi-class change detection.

Dense posterior discrepancy replaces explicit instance matching in CoRegOVCD. After Competitive Posterior Calibration (CPC), it defines the Semantic Posterior Delta (SPD) as

$x_{text}$1

The discrepancy is then filtered by Geometry-Token Consistency Gate (GeoGate) using Depth Anything 3, and regionally regularized by Regional Consensus Discrepancy (RCD) with SLIC superpixels (Tang et al., 2 Apr 2026). This shifts change inference from discrete object correspondence to comparison of competition-aware queried-concept posteriors.

Pixel-wise parsing plus reliability reasoning is the logic of ReA-OVCD. It begins with hard label discrepancy,

$x_{text}$2

then evaluates semantic reliability through Jensen-Shannon divergence and maximum per-class logit difference, and finally validates candidate regions spatially by measuring whether they contain reliable interior pixels far from unstable boundaries (Zhu et al., 18 Jun 2026). The paper explicitly positions this as a response to the trade-off between coarse but stable instance matching and detailed but noisy direct pixel comparison.

Temporal propagation and memory reasoning are most explicit in MemOVCD. Coarse query-specific masks initialize SAM 3 memory, mask evidence is propagated bidirectionally over histogram-aligned transition frames, stable regions are pooled into a query-specific visual exemplar, and patch-wise and global-view predictions are fused by connected-component-aware global-local rectification (Kuang et al., 29 Apr 2026). This is one of the clearest departures from OVCD as static differencing.

Change-specific adapters appear in Seg2Change. Its Feature Modulation Module (FMM), Bi-temporal Difference Fusion Module (BDFM), Effective Difference Query Attention (EDQA), and Residual Upsampler (ResUp) learn a category-agnostic change map from CA-CDD, after which semantic segmentation maps index changed pixels to classes (Su et al., 13 Apr 2026). This avoids proposal generation and fixed similarity thresholds.

5. Benchmarks and empirical landscape

OVCD is evaluated on a recurring set of public benchmarks, but protocols vary substantially across papers.

Benchmark	Typical use in OVCD papers	Notes appearing in the literature
LEVIR-CD	Building change detection	637 pairs total in UniVCD; 128 test pairs
WHU-CD / WHU-CD-256	Building change detection	Patch and original-image evaluation both appear
SECOND	Semantic or multi-class change detection	Six classes: building, tree, water, low vegetation, non-vegetated surface, playground
DSIFN	Land-cover or building-related change	Used in AdaptOVCD, Seg2Change, ReA-OVCD
S2Looking	Building change detection	Used in DynamicEarth, OmniOVCD, MemOVCD
BANDON	Building change detection	Used in DynamicEarth and MemOVCD
Hi-UCD mini	Error attribution / semantic change detection	Used in OpenDPR
SC-SCD / CLCD	Semantic or land-cover change detection	Used in Seg2Change

DynamicEarth established an initial training-free baseline regime. Its best building result in the reported table was SAM-DINOv2-SegEarth-OV with 36.6 IoU / 53.6 F1 on LEVIR-CD and 40.6 IoU / 57.7 F1 on WHU-CD; it also reported that OVCD methods beat the best cross-dataset supervised result on all test sets by +7.9 IoU / +7.0 F1 on LEVIR-CD, +10.8 IoU / +8.9 F1 on WHU-CD, +11.9 IoU / +17.1 F1 on S2Looking, and +13.8 IoU / +22.8 F1 on BANDON (Li et al., 22 Jan 2025).

Later methods improved markedly but under different settings. UniVCD reported UniVCD (orig., postproc) with F1 = 70.7, IoU = 54.7, mIoU = 75.6 on LEVIR-CD and F1 = 76.5, IoU = 61.9, mIoU = 79.6 on WHU-CD, while highlighting strong category-wise performance on SECOND, especially building: F1 58.4, IoU 41.2 and non-vegetated surface: F1 42.6, IoU 27.1 (Zhu et al., 15 Dec 2025). AdaptOVCD reported $x_{text}$3 values of 68.00% on LEVIR-CD, 76.53% on WHU-CD, 59.47% on DSIFN, and 63.81% on Building on SECOND, and stated that it reaches 84.89% of the fully-supervised performance upper bound in cross-dataset evaluation (Dou et al., 6 Feb 2026).

OmniOVCD reported 67.2 IoU / 80.4 F1 on LEVIR-CD, 66.5 IoU / 79.9 F1 on WHU-CD, 24.5 IoU / 39.4 F1 on S2Looking, and a SECOND class average of 27.1 / 41.8 (Zhang et al., 20 Jan 2026). CoRegOVCD emphasized relative gains over the strongest previous training-free baseline, reporting +2.61 F1$x_{text}$4 on LEVIR-CD, +2.24 F1$x_{text}$5 on WHU-CD-256, +4.98 F1$x_{text}$6 on DSIFN, and +1.88 F1$x_{text}$7 on SECOND-Building, with a six-class SECOND average of 47.50% F1$x_{text}$8 and 31.67% IoU$x_{text}$9 (Tang et al., 2 Apr 2026). MemOVCD reported 72.5 $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$0 and 84.1 $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$1 on LEVIR-CD, 26.0 $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$2 and 41.3 $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$3 on S2Looking, 23.2 $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$4 and 37.7 $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$5 on BANDON, and 37.8 $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$6 and 54.9 $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$7 on DSIFN (Kuang et al., 29 Apr 2026).

Other strands reached strong results under their own protocols. Seg2Change reported WHU-CD: $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$8, $$\mathcal{M}: (I_a, I_b, \mathcal{T}) \rightarrow M \in \{0,1\}^{H \times W},$$9, LEVIR-CD: $I_a$0, $I_a$1, and SECOND: $I_a$2, $I_a$3 (Su et al., 13 Apr 2026). ReA-OVCD reported LEVIR-CD: $I_a$4, $I_a$5, WHU-CD: $I_a$6, $I_a$7, DSIFN: $I_a$8, $I_a$9, and a SECOND class-average of 47.19% $I_b$0 and 32.02% IoU$I_b$1 (Zhu et al., 18 Jun 2026). OpenDPR, focused on the recognition bottleneck, reported 37.3 mIoU / 54.3 mF1 in SECOND multi-class inference and OpenDPR-W average mIoU 19.7 and mF1 30.4 on Hi-UCD mini (Guo et al., 29 Mar 2026).

Because some methods are training-free, some are unsupervised, some are weakly supervised, and some train adapters or heads, direct numerical comparison across papers is not always protocol-equivalent. This suggests that the empirical landscape is best read as evidence of rapid progress under multiple supervision and inference regimes rather than as a single strictly ordered leaderboard.

6. Limitations, controversies, and research directions

Several limitations recur across the OVCD literature. The first is dataset and benchmark scarcity. Semantic-CD explicitly states that available SCD datasets are too small in category diversity for a fully standard OVCD benchmark (Zhu et al., 12 Jan 2025), while DynamicEarth emphasizes that existing change detection datasets are limited, scattered, and heterogeneous, with poor annotation quality and no standardized open-vocabulary labels (Li et al., 22 Jan 2025). Seg2Change’s construction of CA-CDD can be read as a response to this gap, but it also indicates that broader category-agnostic supervision remains a bottleneck (Su et al., 13 Apr 2026).

The second is disagreement over the primary technical bottleneck. OpenDPR argues that category identification is the main bottleneck, not localization, because CLIP-style image-text matching compresses fine-grained land-cover semantics too aggressively for remote sensing (Guo et al., 29 Mar 2026). ReA-OVCD instead frames the problem as a trade-off between instance-level comparison, which misses fine-grained or partial changes, and direct pixel comparison, which is unstable and boundary-noisy (Zhu et al., 18 Jun 2026). MemOVCD criticizes weak temporal coupling in late-comparison systems (Kuang et al., 29 Apr 2026). These are not mutually exclusive diagnoses, but they emphasize different points of failure.

The third limitation is domain sensitivity. Multiple papers note that remote sensing differs from natural-image pretraining in viewpoint, scale variation, orientation ambiguity, radiometric variation, and land-cover subtlety (Dou et al., 6 Feb 2026, Guo et al., 29 Mar 2026). Concrete manifestations are category-specific failures. UniVCD reports weaker performance on water with standard CLIP and notes that replacing CLIP with RemoteCLIP improves the water category substantially (Zhu et al., 15 Dec 2025). AdaptOVCD lists Tree and Low Vegetation as harder classes due to semantic ambiguity and weak boundaries (Dou et al., 6 Feb 2026). ReA-OVCD notes weaker performance on Playground because of sample scarcity and semantic ambiguity (Zhu et al., 18 Jun 2026).

The fourth limitation is post-processing and threshold sensitivity. UniVCD’s post-processing can improve precision and IoU but may reduce recall in ambiguous scenes (Zhu et al., 15 Dec 2025). AdaptOVCD states that prompts and thresholds need scene-specific tuning (Dou et al., 6 Feb 2026). MemOVCD reports that moderate histogram-aligned bridging helps substantially, with $I_b$2 performing best, while $I_b$3 begins to degrade (Kuang et al., 29 Apr 2026). CoRegOVCD and ReA-OVCD both depend on calibrated dense scores and structural filters, underscoring that fully training-free inference often shifts complexity from optimization to calibration and rule design (Tang et al., 2 Apr 2026, Zhu et al., 18 Jun 2026).

Research directions identified in the papers are comparatively consistent. They include stronger remote sensing-adapted foundation models, improved prompt engineering or LLM-generated prompt vocabularies, better handling of multi-class inference, more reliable treatment of small objects and internal-part changes, and more diverse datasets with richer open-vocabulary annotations (Li et al., 22 Jan 2025, Dou et al., 6 Feb 2026). Another clear trajectory is toward greater temporal coupling and internal model coherence: OmniOVCD moves to a standalone SAM 3 design (Zhang et al., 20 Jan 2026), MemOVCD turns OVCD into memory-based tracking (Kuang et al., 29 Apr 2026), and CoRegOVCD and ReA-OVCD replace brittle direct differencing with calibrated posterior or reliability-aware reasoning (Tang et al., 2 Apr 2026, Zhu et al., 18 Jun 2026). This suggests a broader consolidation of OVCD around semantically calibrated, temporally aware, and structurally validated inference rather than around naïve before/after comparison alone.