OpenDPR: Open-Vocabulary Change Detection via Vision-Centric Diffusion-Guided Prototype Retrieval for Remote Sensing Imagery

Qi Guo; Jue Wang; Yinhe Liu; Yanfei Zhong

cs.CVMar 29, 2026

OpenDPR: Open-Vocabulary Change Detection via Vision-Centric Diffusion-Guided Prototype Retrieval for Remote Sensing Imagery

Authors: Qi Guo, Jue Wang, Yinhe Liu, Yanfei Zhong

Organizations: Wuhan University · Beijing Institute of Technology

Abstract

Open-vocabulary change detection (OVCD) seeks to recognize arbitrary changes of interest by enabling generalization beyond a fixed set of predefined classes. We reformulate OVCD as a two-stage pipeline: first generate class-agnostic change proposals using visual foundation models (VFMs) such as SAM and DINOv2, and then perform category identification with vision-language models (VLMs) such as CLIP. We reveal that category identification errors are the primary bottleneck of OVCD, mainly due to the limited ability of VLMs based on image-text matching to represent fine-grained land-cover categories. To address this, we propose OpenDPR, a training-free vision-centric diffusion-guided prototype retrieval framework. OpenDPR leverages diffusion models to construct diverse prototypes for target categories offline, and to perform similarity retrieval with change proposals in the visual space during inference. The secondary bottleneck lies in change localization, due to the inherent lack of change priors in VFMs. To bridge this gap, we design a spatial-to-change weakly supervised change detection module named S2C to adapt their strong spatial modeling capabilities for change localization. Integrating the pretrained S2C into OpenDPR leads to an optional weakly supervised variant named OpenDPR-W, which further improves OVCD with minimal supervision. Experimental results on four benchmark datasets demonstrate that the proposed methods achieve state-of-the-art performance under both supervision modes. Code is available at https://github.com/guoqi2002/OpenDPR.

Figures & tables

Figure 1: (a) Our analysis reveals that category identification errors are the primary bottleneck in OVCD. (b) OpenDPR performs prototype retrieval with change proposals in the visual space to enhance category identification, while OpenDPR-W further improves change localization with additional weak supervision.

Figure 2: Overview of the proposed training-free OpenDPR and its weakly supervised variant OpenDPR-W. OpenDPR constructs a semantically diverse prototype set offline and identifies categories by retrieving change proposals against prototypes in the visual space during inference. OpenDPR-W further improves change localization by integrating the weakly supervised S2C module.

Figure 3: Overview of the proposed S2C, which refines CAM-based change pseudo-labels using spatial priors from VFMs and adapts them to binary change localization through pre-training.

Method	LEVIR-CD		WHU-CD
Method	IoU	F1	IoU	F1
CVA [ 2 ]	—	12.2	—	—
PCA-KM [ 5 ]	4.8	9.1	5.4	10.2
CNN-CD [ 12 ]	7.0	13.1	4.9	9.4
DSFA [ 11 ]	4.3	8.2	4.1	7.8
DCVA [ 31 ]	7.6	14.1	10.9	19.6

Table 1: Comparison of unsupervised methods on building change detection datasets. “—” denotes data missing. Best and second-best scores are marked in bold and in underline , respectively.

Method	LEVIR-CD		WHU-CD		Params
Method	IoU	F1	IoU	F1	Params
FCD-GAN [ 41 ]	30.5	43.1	39.3	56.5	43.4M
BGMix [ 14 ]	38.1	52.0	42.7	59.8	22.1M
CS-WSCDNet [ 40 ]	40.7	57.8	57.3	72.8	99.9M
MS-Former [ 22 ]	39.8	57.0	51.3	67.8	15.2M
TransWCD [ 48 ]	42.9	60.1	52.4	68.7	17.9M

Table 2: Comparison of weakly supervised methods on building change detection datasets. “Params” denotes trainable parameters.

Method	Water		Grass		Building		Greenhouse		Road		Bridge		Bareland		Woodland		Average
	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1	mIoU	mF1
Per-class Inference
I-M-C [ 19 ]	10.0	18.1	9.2	16.8	35.1	52.0	—	—	18.6	31.3	—	—	—	—	—	—	9.1	14.8
M-C-I [ 19 ]	2.6	5.0	21.0	34.8	22.7	37.1	8.9	16.3	22.3	36.5	—	—	26.5	41.9	—	—	13.0	21.5
OpenDPR (Ours)	2.9	5.5	26.6	42.1	26.5	41.8	15.1	26.2	18.2	30.8	—	—	24.2	39.0	1.7	3.4	14.4	23.6
OpenDPR-W (Ours)	4.4	8.4	32.6	49.2	34.2	51.0	12.9	22.8	21.6	35.5	—	—	29.0	44.9	3.3	6.4	17.3	27.3

Table 3: Comparison of OVCD methods on the semantic change detection dataset Hi-UCD mini. “—” denotes that the score is close to 0.

Method	Water		Ground		Low vegetation		Tree		Building		Playground		Average
	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1	mIoU	mF1
Per-class Inference
I-M-C [ 19 ]	12.2	21.7	—	—	1.4	2.7	14.1	24.8	28.1	43.9	16.0	27.6	12.0	20.1
M-C-I [ 19 ]	14.3	25.1	26.2	41.6	24.1	38.9	20.3	33.8	38.1	55.2	20.0	33.3	23.8	38.0
OpenDPR (Ours)	12.0	21.4	31.2	47.5	23.9	38.6	19.7	33.0	43.2	60.4	26.5	42.0	26.1	40.5
Multi-class Inference

Table 4: Comparison of OVCD methods on the semantic change detection dataset SECOND. “—” denotes that the score is close to 0.

Figure 4: Qualitative evaluation of the proposed methods on four change detection datasets.

Proto	Encoder	Match	$\mathcal{L}$	$\mathcal{W}$	$\mathcal{H}$	$\mathcal{S}$
LLM	CLIP	mean	37.96	39.06	13.31	25.16
	CLIP	max	40.58	47.40	14.61	26.98
	DINO	mean	31.83	20.55	6.24	21.93
	DINO	max	33.02	26.41	8.17	20.89
	DINOv2	mean	36.65	43.64	16.74	21.27
	DINOv2	max	44.83	54.31	17.52	28.55

Table 5: Ablation study on (mean) IoU of different components in OpenDPR.

\mathcal{L}

\mathcal{W}

\mathcal{H}

, and

\mathcal{S}

denote the LEVIR-CD, WHU-CD, Hi-UCD mini, and SECOND datasets, respectively.

OVCD Method		S2C Components		$\mathcal{L}$	$\mathcal{W}$	$\mathcal{H}$
OVCD Method		Base	Refine	$\mathcal{L}$	$\mathcal{W}$	$\mathcal{H}$
—		✓	×	43.26	61.84	—
—		✓	✓	44.67	64.83	—
M-C-I		×	×	36.63	40.60	13.94
	-W	✓	×	40.74	62.44	14.79
	-W	✓	✓	41.48	64.43	15.85

Table 6: Ablation study on (mean) IoU of S2C and its integration with training-free OVCD models.

\mathcal{L}

\mathcal{W}

, and

\mathcal{H}

denote the LEVIR-CD, WHU-CD, and Hi-UCD mini datasets, respectively.

Figure 5: Prototype sensitivity analysis on OpenDPR and OpenDPR-W. Only OpenDPR is evaluated on SECOND.

Remote Sensing Change Detection

Remote sensing change detection finds and interprets differences across Earth-observation images acquired at different times.

See all 89 papers

ReA-OVCD: Training-Free Open-Vocabulary Change Detection via Semantic-Spatial Reliability Assessment
Jun 18, 2026Hongming Zhu, Huaji Chen, Bowen Du +2Remote Sensing Change Detection Semantic Change Detection
School of Computer Science and Technology, Tongji University, Shanghai 200092, China · College of Surveying and Geo-Informatics, Tongji University, Shanghai 200092, China
Zero-OVCD: Bridging Training-Free Foundation Models and Pseudo-Label Learning for Open-Vocabulary Change Detection
Aug 12, 2026Daifeng Peng, Yuanke Peng, Haiyan GuanRemote Sensing Change Detection Pseudo-Labeling
School of Remote Sensing and Geomatics Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China
Seg2Change: Adapting Open-Vocabulary Semantic Segmentation Model for Remote Sensing Change Detection
Apr 13, 2026You Su, Yonghong Song, Jingqi Chen +1Remote Sensing Change Detection
Xi’an Jiaotong University, Xi’an, Shaanxi, China

Jun 18, 2026J/K move · Enter open · S save

ReA-OVCD: Training-Free Open-Vocabulary Change Detection via Semantic-Spatial Reliability Assessment

Hongming Zhu, Huaji Chen, Bowen Du, Sicong Liu+1

School of Computer Science and Technology, Tongji University, Shanghai 200092, China · College of Surveying and Geo-Informatics, Tongji University, Shanghai 200092, China

Remote Sensing Change Detection Semantic Change Detection

Open paper AskPDF arXiv

Unlike traditional remote sensing change detection that relies on predefined categories, Open-Vocabulary Change Detection (OVCD) identifies land cover changes flexibly using arbitrary text prompts. However, most existing OVCD methods rely on instance-level matching for stable correspondence but may overlook fine-grained variations (e.g., partial building extensions). Dense pixel-level comparison is more flexible, yet direct semantic comparison often produces unreliable candidate changes due to semantic ambiguity and spatial inconsistency. To this end, we propose ReA-OVCD, an efficient training-free framework that revisits pixel-level OVCD from a reliability assessment perspective. It first derives candidate change regions from pixel-wise semantic discrepancies to retain flexible localization. Instead of directly trusting these candidates, ReA-OVCD applies a two-stage semantic-spatial reliability assessment. The semantic stage evaluates whether a label discrepancy is supported by meaningful distributional and response-level changes, while the spatial stage validates whether a candidate region contains stable interior evidence rather than only boundary-induced responses. Extensive experiments across LEVIR-CD, WHU-CD, DSIFN, and SECOND show that the proposed framework improves the reliability of pixel-level OVCD and consistently outperforms state-of-the-art approaches, achieving

\mathrm{F}_{1}^{C}

improvements of 3.54% to 8.45% while maintaining superior computational efficiency. The code is available at https://github.com/Funny0101/ReA-OVCD.

Abstract

Remote Sensing Change Detection

Explore similar work

ReA-OVCD: Training-Free Open-Vocabulary Change Detection via Semantic-Spatial Reliability Assessment

Zero-OVCD: Bridging Training-Free Foundation Models and Pseudo-Label Learning for Open-Vocabulary Change Detection

Seg2Change: Adapting Open-Vocabulary Semantic Segmentation Model for Remote Sensing Change Detection