Research map / Detection and segmentation / Open-set and few-shot detection
Segment Any Change
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AnyChange, a training-free adaptation of SAM for zero-shot change detection in bitemporal remote sensing images via bitemporal latent matching, plus a point query mechanism for object-centric change.
Problem
Zero-shot change detection, i.e., detecting class-agnostic changes in bitemporal images on unseen change types and data distributions, has not been explored; existing change detectors need training on change data and do not generalize.
Approach
Exploits intra-image and inter-image semantic similarity in SAM's latent space: mask embeddings (averaged image embeddings over SAM proposals) are matched bidirectionally across two dates using negative cosine similarity, with no training; a point query averages embeddings of clicked objects to filter for object-centric changes.
Key results
- On SECOND unsupervised change detection, AnyChange (ViT-B) with ChangeStar pseudo-label training reaches 48.2% F1, beating I3PE (43.8%).
- Zero-shot AnyChange (ViT-H) reaches 41.8% F1 on SECOND, exceeding DINOv2+CVA's up-to-3.2% margin claim context and other baselines in pixel-level change recall.
- Supervised S2Looking with ChangeStar trained on 3,500 AnyChange pseudo-labels (single point per pair) achieves 40.2% F1 versus 66.3% with full ground truth.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace mask IoU matching with bitemporal latent matching | SAM-based zero-shot change proposals · LEVIR-CD | mask AR@1000: 9.0 -> 32.6 | Helped |
| replace mask IoU matching with bitemporal latent matching | SAM-based zero-shot change proposals · S2Looking (binary) | mask AR@1000: 31.5 -> 48.3 | Helped |
| replace pixel-wise CVA matching with instance-wise bitemporal latent matching | SAM-based zero-shot change proposals · xView2 (binary) | mask AR@1000: 12.2 -> 27.8 | Helped |
| use bitemporal latent matching instead of DINOv2+CVA | zero-shot change detection · SECOND (binary) | F1: 41.4 -> 44.6 | Helped |
| train ChangeStar on AnyChange ViT-B pseudo-labels | ChangeStar (1×256) · SECOND test | F1: 43.8 -> 48.2 | Helped |
| train ChangeStar with AnyChange pseudo-labels instead of 1% ground truth | ChangeStar (1×96) · S2Looking test | F1: 37.2 -> 40.2 | Helped |
| make matching bidirectional rather than one-way from t to t+1 (ablation) | AnyChange · LEVIR-CD | mask AR@1000: 1.3 -> 32.6 | Helped |
| make matching bidirectional rather than one-way from t+1 to t (ablation) | AnyChange · LEVIR-CD | mask AR@1000: 35.9 -> 32.6 | Hurt |
| apply independent random color jitter to the bitemporal images (ablation) | AnyChange · LEVIR-CD | mask AR@1000: 32.6 -> 30.7 | Hurt |
| apply independent random color jitter to the bitemporal images (ablation) | AnyChange · S2Looking (binary) | mask AR@1000: 48.3 -> 48.4 | No clear effect |
| apply independent random color jitter to the bitemporal images (ablation) | AnyChange · xView2 (binary) | mask AR@1000: 27.8 -> 25.9 | Hurt |
| apply independent random color jitter to the bitemporal images (ablation) | AnyChange · SECOND (binary) | mask AR@1000: 27.0 -> 26.4 | Hurt |
Surprising findings
- Training on pseudo-labels from AnyChange yields higher performance than the pseudo-labels themselves, suggesting the model learns dataset-specific biases.
- Single-directional latent matching is highly sensitive to temporal order (e.g., mask AR on LEVIR-CD of 1.3% vs 35.9%).
Limitations
- Zero-shot change detection evaluation lacks a concept-complete class-agnostic change dataset, so precision and F1 on building-centric datasets are only approximate.
- Inherits SAM's latent-space biases and is sensitive to radiation variation (mask AR jitter up to about 1.9%).
- Less effective for tiny or minor object changes unless the image is upsampled.
Experimental setup
Datasets: LEVIR-CD S2Looking xView2 SECOND
Benchmarks: LEVIR-CD S2Looking xView2 SECOND
Metrics: F1 precision recall mask AR@1000
Baselines: CVA DINOv2+CVA SAM+Mask Match SAM+CVA Match I3PE ISFA DSFA DCAE OBCD IRMAD KPCA-MNet DCVA
Builds on: Segment Anything Model DINOv2 change vector analysis ChangeStar LoRA
Closest papers by meaning
- Towards Generalizable Scene Change Detection (CVPR 2025 · similarity 0.89)
- Zero-Shot Scene Change Detection (AAAI 2025 · similarity 0.89)
- Change3D: Revisiting Change Detection and Captioning from A Video Modeling Perspective (CVPR 2025 · similarity 0.86)
- MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning (NeurIPS 2024 · similarity 0.85)
- Improving the Generalization of Segmentation Foundation Model under Distribution Shift via Weakly Supervised Adaptation (CVPR 2024 · similarity 0.85)
- Uncertainty-aware Fine-tuning of Segmentation Foundation Models (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.