Organizations: China Agricultural University, College of Information and Electrical Engineering, Beijing, China · China Agricultural University, College of Science, Beijing, China
Remote sensing change detection (CD) aims to identify where land-cover semantics change across time, but most existing methods still assume a fixed label space and therefore cannot answer arbitrary user-defined queries. Open-vocabulary change detection (OVCD) instead asks for the change mask of a queried concept. In the fully training-free setting, however, dense concept responses are difficult to compare directly across dates: appearance variation, weak cross-concept competition, and the spatial continuity of many land-cover categories often produce noisy, fragmented, and semantically unreliable change evidence. We propose Consistency-Regularized Open-Vocabulary Change Detection (CoRegOVCD), a training-free dense inference framework that reformulates concept-specific change as calibrated posterior discrepancy. Competitive Posterior Calibration (CPC) and the Semantic Posterior Delta (SPD) convert raw concept responses into competition-aware queried-concept posteriors and quantify their cross-temporal discrepancy, making semantic change evidence more comparable without explicit instance matching. Geometry-Token Consistency Gate (GeoGate) and Regional Consensus Discrepancy (RCD) further suppress unsupported responses and improve spatial coherence through geometry-aware structural verification and regional consensus. Across four benchmarks spanning building-oriented and multi-class settings, CoRegOVCD consistently improves over the strongest previous training-free baseline by 2.24 to 4.98 F1C points and reaches a six-class average of 47.50% F1C on SECOND.
Figures & tables
Figure 1. Paradigm comparison between previous training-free OVCD methods and CoRegOVCD. (a) Previous methods typically rely on explicit mask or instance representations, semantic cues, and cross-temporal matching before producing the final change mask. (b) CoRegOVCD instead performs dense posterior-based change inference through dense score construction, CPC, SPD, GeoGate, RCD, and lightweight final mask inference. A two-part flowchart comparing prior training-free OVCD paradigms with CoRegOVCD. In panel (a), a text query and two temporal images feed a multi-stage pipeline with hard mask maps, semantic cues, cross-temporal matching, and change integration, followed by a change mask. In panel (b), the same inputs feed a dense score, CPC, SPD, GeoGate, RCD, and final mask inference pipeline, which then outputs the change mask.
Figure 2. Overview of CoRegOVCD. Given bi-temporal images and a queried concept, the framework first constructs dense concept confidence scores from prompt-conditioned Segment Anything Model 3 (SAM 3) outputs and calibrates them with Competitive Posterior Calibration (CPC). It then computes the Semantic Posterior Delta (SPD) as the semantic change signal. For structural verification, Geometry-Token Consistency Gate (GeoGate) employs a Geometric Encoder instantiated with Depth Anything 3 (DA3) to extract geometry tokens and derive the gate map G . Finally, Regional Consensus Discrepancy (RCD) fuses SPD and the gate map G and imposes SLIC-based regional consensus, after which a lightweight final mask inference stage converts the resulting score map into the final change mask. Overview diagram of CoRegOVCD. The left side shows two temporal input images and a text query. The center shows dense score construction and CPC on prompt-conditioned SAM~3 outputs, SPD computation, GeoGate built from the Geometric Encoder instantiated with Depth Anything 3 and the derived gate map, RCD with SLIC-based regional pooling, and a final mask inference stage with thresholding and structured filtering. The right side shows the final change mask.
LEVIR-CD (%)
WHU-CD-256 (%)
DSIFN (%)
SECOND (Building) (%)
Method
Prec.
Rec.
IoU C
F1 C
Prec.
Rec.
IoU C
F1 C
Prec.
Rec.
IoU C
F1 C
Prec.
Rec.
IoU C
F1 C
Traditional unsupervised methods
CVA ( Bovolo and Bruzzone, 2007 )
5.43
93.93
5.41
10.26
3.73
92.20
3.71
7.16
18.02
92.16
17.75
30.14
11.65
97.08
11.61
20.81
IRMAD ( Nielsen, 2007 )
11.06
14.62
6.72
12.59
4.19
13.64
3.31
6.41
28.63
7.05
5.99
11.31
23.26
22.60
12.95
22.93
PCA-Kmeans ( Celik, 2009 )
5.92
36.13
5.36
10.18
7.24
44.72
6.64
12.46
27.28
43.16
20.07
33.43
17.74
44.32
14.51
25.34
ISFA ( Wu et al., 2014 )
5.37
96.27
5.36
10.17
3.83
92.99
3.82
7.36
17.51
96.81
17.41
29.66
11.46
98.49
11.44
20.52
Table 1. Quantitative comparison on three building-oriented benchmarks and SECOND-Building under the per-class OR-rule evaluation protocol. For SECOND-Building, a pixel is positive if the queried class appears at either timestamp. “-” denotes values not numerically reported in the corresponding paper. The best and second-best results are highlighted. F1 C serves as the primary evaluation metric.
Building (%)
Tree (%)
Water (%)
Low vegetation (%)
Surface (%)
Playground (%)
Class Avg. (%)
Method
IoU C
F1 C
IoU C
F1 C
IoU C
F1 C
IoU C
F1 C
IoU C
F1 C
IoU C
F1 C
IoU C
F1 C
DynamicEarth (IMC) ( Li et al., 2025a )
26.50
42.00
13.50
23.80
9.80
17.90
0.00
0.00
0.00
0.00
16.50
28.30
11.05
18.67
DynamicEarth (MCI) ( Li et al., 2025a )
38.10
55.20
20.30
33.80
14.30
25.10
24.10
38.90
26.20
41.60
20.00
33.30
23.83
37.98
AdaptOVCD ( Dou et al., 2026 )
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
OmniOVCD ( Zhang et al., 2026 )
45.20
62.30
16.70
28.60
21.20
35.00
24.50
39.30
27.70
43.40
27.00
42.40
27.10
41.80
CoRegOVCD (Ours)
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
Table 2. Open-vocabulary change detection results on the six SECOND classes under the per-class OR-rule evaluation, where a pixel is positive for class c if either timestamp belongs to c . The best and second-best results are highlighted. F1 C serves as the primary evaluation metric.
DSIFN (%)
LEVIR-CD (%)
SECOND Avg. (%)
Setting
IoU C
F1 C
IoU C
F1 C
IoU C
F1 C
w/o CPC
45.66
62.70
70.30
82.56
31.05
46.65
w/o GeoGate
34.51
51.31
50.01
66.67
22.40
36.21
RCD ( w/o additive term)
23.26
37.74
69.92
82.05
31.16
46.85
RCD ( w/o SLIC)
44.85
61.93
68.89
81.32
29.60
45.36
Decoding ( w/o StructFilter)
42.32
59.47
70.15
82.46
31.12
46.81
Table 3. Ablation of CPC, GeoGate, RCD, and final mask inference in CoRegOVCD. “RCD ( w/o additive term)” removes the additive compensation term in Eq. ( 12 ), and “Decoding ( w/o StructFilter)” skips the structural filtering step in final mask inference. SECOND Avg. reports the class-average result over the six semantic categories.
DSIFN (%)
LEVIR-CD (%)
SECOND Avg. (%)
Gate
IoU C
F1 C
IoU C
F1 C
IoU C
F1 C
None
34.51
51.31
50.01
66.67
22.40
36.21
Depth
19.58
32.75
46.12
63.13
22.12
35.29
Hybrid
25.60
40.76
66.89
80.16
26.58
40.72
GeoGate
47.55
64.45
70.95
83.01
31.67
47.50
Table 4. Comparison of GeoGate designs. “Depth” uses only a depth-based gate from DA3, whereas “Hybrid” combines the depth-based gate with the geometry-token consistency gate. SECOND Avg. reports the class-average result over the six semantic categories.
Figure 3. Open-vocabulary query substitution on SECOND. Each panel reports F1 C (%) for semantically related query words of one class. Bold bars denote the default query, and the hatched bar denotes a prompt set. fig_S3_prompt_substitution
Figure 4. Efficiency and accuracy trade-off among training-free OVCD methods. (a) Latency per image pair and peak memory on an NVIDIA A800-SXM4-80GB device, measured after a one-pair warm-up. (b) Accuracy–speed trade-off in terms of throughput and the corresponding DSIFN F1 C values used in the efficiency comparison. A two-panel figure comparing efficiency and accuracy of DynamicEarth MCI, DynamicEarth IMC, AdaptOVCD, and CoRegOVCD. The left panel shows bars for latency and a line for peak memory. The right panel shows a scatter plot of throughput versus DSIFN F1_C, where CoRegOVCD lies in the top-right region with both the highest throughput and highest accuracy.
Figure 5. Qualitative comparison on SECOND with DynamicEarth (MCI/IMC), AdaptOVCD, and CoRegOVCD across six semantic categories. CoRegOVCD yields more complete masks with fewer unsupported responses. Single-column qualitative comparison figure on SECOND. The columns show temporal images, ground truth, predictions from DynamicEarth MCI, DynamicEarth IMC, AdaptOVCD, and CoRegOVCD across six semantic categories.
Figure 6. Intermediate responses and final predictions of CoRegOVCD on three representative samples. GeoGate suppresses unsupported responses, and RCD produces cleaner, more localized change evidence before final decoding. Single-column qualitative figure showing three representative examples with temporal images, ground truth, final prediction, the SPD map, the GeoGate output, and the RCD score map. The intermediate maps show how GeoGate suppresses noisy appearance changes before the final mask inference stage converts the RCD output into the final mask.
School of Computer Science and Technology, Tongji University, Shanghai 200092, China · College of Surveying and Geo-Informatics, Tongji University, Shanghai 200092, China