Segment Any Change

TL;DR

This work introduces AnyChange, a zero-shot change detection framework built on Segment Anything Model (SAM) that achieves training-free adaptation to bitemporal remote sensing images through bidirectional latent matching. By exploiting intra-image and inter-image semantic similarities in SAM's latent space and adding a point-query mechanism, AnyChange can generate both pixel-level and instance-level change masks without change-specific labeled data, while remaining promptable. The approach is evaluated across multiple datasets, achieving state-of-the-art unsupervised performance on SECOND and competitive results with large backbones on zero-shot change tasks, and demonstrates practical utility as a change data engine via pseudo-labeling for supervised and unsupervised learning. Overall, AnyChange establishes a foundation-model path for Earth vision by delivering zero-shot change detection, interactive object-centric changes, and training-free deployment with strong performance and broad applicability in geoscience and environmental monitoring.

Abstract

Visual foundation models have achieved remarkable results in zero-shot image classification and segmentation, but zero-shot change detection remains an open problem. In this paper, we propose the segment any change models (AnyChange), a new type of change detection model that supports zero-shot prediction and generalization on unseen change types and data distributions. AnyChange is built on the segment anything model (SAM) via our training-free adaptation method, bitemporal latent matching. By revealing and exploiting intra-image and inter-image semantic similarities in SAM's latent space, bitemporal latent matching endows SAM with zero-shot change detection capabilities in a training-free way. We also propose a point query mechanism to enable AnyChange's zero-shot object-centric change detection capability. We perform extensive experiments to confirm the effectiveness of AnyChange for zero-shot change detection. AnyChange sets a new record on the SECOND benchmark for unsupervised change detection, exceeding the previous SOTA by up to 4.4% F$_1$ score, and achieving comparable accuracy with negligible manual annotations (1 pixel per image) for supervised change detection. Code is available at https://github.com/Z-Zheng/pytorch-change-models.

Figures (8)

• Figure 1: Zero-Shot Change Detection with AnyChange on a wide range of application scenarios in geoscience. Each subfigure presents the pre-event image, the post-event image, and their change instance masks in order. The boundary of each change instance mask is rendered by cyan, and meanwhile, these change masks are also drawn on pre/post-event images to show more clearly where the change occurred. The color of each change mask is used to distinguish between different instances.
• Figure 2: Segment Any Change Models, AnyChange. SAM forward: given grid points $\{\mathbf{p}_i\}$ as prompts and input images, SAM produces object masks $\{\mathbf{m}_{t,i}\}$ and image embedding $\mathbf{z_t}$ on the image at time $t$. Bitemporal Latent Matching does a bidirectional matching to compute the change confidence score for each change proposal, and then top-k sorting or thresholding is applied for zero-shot change proposal and detection. Point Query allows users to click some points (the case of two points in this subfigure) with the same category to filter class-agnostic change masks via semantic similarity for object-centric change detection.
• Figure 3: intra-image semantic similarity
• Figure 4: inter-image semantic similarity
• Figure 6: Examples of Point Query Mechanism. The effects of w/o point query, one-point query, and three-point queries are shown in sequence from left to right. (best viewed digitally with zoom, especially for the red points)