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

Abstract

Open-vocabulary change detection (OVCD) aims to identify arbitrary changes of interest by extending generalization capabilities beyond a restricted set of predefined categories. This study reimagines OVCD as a two-phase workflow: initially, class-agnostic change proposals are generated using visual foundation models (VFMs) like DINOv2 and SAM; subsequently, category identification is executed via vision-language models (VLMs) such as CLIP. Our analysis indicates that the primary limitation of OVCD stems from errors in category identification, largely attributable to the constrained capacity of image-text matching-based VLMs to accurately represent fine-grained land-cover types.

To overcome this hurdle, we introduce OpenDPR, a training-free framework that employs vision-centric, diffusion-guided prototype retrieval. OpenDPR utilizes diffusion models to generate diverse prototypes for target categories during an offline phase, facilitating similarity retrieval against change proposals within the visual space during inference. A secondary challenge involves change localization, which is hindered by the inherent absence of change priors in VFMs. To resolve this, we developed S2C, a spatial-to-change weakly supervised change detection module designed to harness the robust spatial modeling strengths of VFMs for precise change localization. By integrating the pretrained S2C into OpenDPR, we create an optional weakly supervised variant, OpenDPR-W, which enhances OVCD performance with minimal supervisory input. Evaluations across four benchmark datasets confirm that our proposed methods deliver state-of-the-art results under both supervision paradigms. The codebase is accessible at https://github.com/guoqi2002/OpenDPR.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC