Research map / Detection and segmentation / Open-set and few-shot detection
Towards Generalizable Scene Change Detection
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 GeSCF, a training-free zero-shot scene change detector built on SAM's encoder features plus mask matching, along with the GeSCD benchmark, TC metric, and the ChangeVPR dataset.
Problem
Scene change detection models trained on narrow datasets fail under unseen environments and reversed temporal order, and produce inconsistent masks when input order is swapped.
Approach
Zero-shot SCD by intercepting SAM ViT key-facet correlations between two images to build an adaptive (skewness-based) pseudo-mask, then refining it with SAM class-agnostic masks via geometric intersection matching and semantic embedding similarity; the design is commutative, so temporal consistency is complete by construction.
Key results
- GeSCF achieves 67.7% average F1 across standard SCD datasets with TC=1.0 everywhere, versus 29.3% for the best prior baseline (C-3PO, uni-temporal).
- On ChangeVPR, GeSCF averages 64.1% F1 versus 24.9% for the best baseline, a gap of about 30 points.
- Zero-shot GeSCF reaches 48.0% F1 on the SECOND remote sensing benchmark, outperforming listed unsupervised methods (best baseline 36.8%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GeSCF instead of the strongest reported baseline | SCD models evaluated on standard SCD datasets · Standard SCD datasets (Avg.) | F1-score: +19.2% | Helped |
| use GeSCF instead of the strongest reported baseline | SCD models evaluated on ChangeVPR · ChangeVPR (Avg.) | F1-score: +30.0% | Helped |
| use GeSCF instead of the strongest reported baseline | unsupervised remote sensing change detection on SECOND · SECOND test | F1-score | Helped |
| add adaptive thresholding instead of a fixed threshold (ablation) | GeSCF pseudo-mask generation · GeSCD evaluation (Table 3) | F1-score: 43.0 -> 53.9 | Helped |
| add adaptive thresholding instead of a fixed threshold (ablation) | GeSCF pseudo-mask generation · GeSCD evaluation (Table 3) | mIoU: 26.3 -> 34.9 | Helped |
| add geometric intersection matching (GIM) (ablation) | GeSCF with adaptive pseudo-masks · GeSCD evaluation (Table 3) | F1-score: 53.9 -> 65.0 | Helped |
| add geometric intersection matching (GIM) (ablation) | GeSCF with adaptive pseudo-masks · GeSCD evaluation (Table 3) | mIoU: 34.9 -> 47.6 | Helped |
| add semantic similarity matching (SSM) (ablation) | GeSCF with pseudo-masks and GIM · GeSCD evaluation (Table 3) | F1-score: 65.0 -> 65.9 | Helped |
| add semantic similarity matching (SSM) (ablation) | GeSCF with pseudo-masks and GIM · GeSCD evaluation (Table 3) | mIoU: 47.6 -> 48.8 | Helped |
| fine-tune correlation layers on relatively small SCD datasets instead of using GeSCF's zero-shot approach (ablation) | SAM-based SCD · GeSCD evaluation (Table 3) | F1-score | Hurt |
| fine-tune CHVA on relatively small SCD datasets instead of using GeSCF's zero-shot approach (ablation) | SAM-based SCD · GeSCD evaluation (Table 3) | F1-score | Hurt |
| fine-tune MTF and MSF on relatively small SCD datasets instead of using GeSCF's zero-shot approach (ablation) | SAM-based SCD · GeSCD evaluation (Table 3) | F1-score | Hurt |
Surprising findings
- Simple bi-temporal training does not guarantee temporal consistency and can worsen unseen-domain and uni-temporal performance.
- Fine-tuning SAM adapters on small SCD datasets undermines its zero-shot capabilities relative to the training-free pseudo-masks.
Limitations
- Relies on a coarse geometric alignment step before feature correlation.
- Some in-domain samples show lower quantitative scores than baselines despite more reasonable masks.
Experimental setup
Datasets: VL-CMU-CD TSUNAMI ChangeSim SECOND
Benchmarks: VL-CMU-CD TSUNAMI ChangeSim ChangeVPR SECOND
Metrics: F1-score mIoU Temporal Consistency Precision Recall
Baselines: CSCDNet CDResNet DR-TANet C-3PO ISFA DSFA DCAE OBCD KPCA-MNet DCVA
Builds on: Segment Anything Model CSCDNet C-3PO DR-TANet
Closest papers by meaning
- Segment Any Change (NeurIPS 2024 · similarity 0.89)
- Zero-Shot Scene Change Detection (AAAI 2025 · similarity 0.88)
- EMPLACE: Self-Supervised Urban Scene Change Detection (AAAI 2025 · similarity 0.88)
- Multi-View Pose-Agnostic Change Localization with Zero Labels (CVPR 2025 · similarity 0.87)
- The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.86)
- Change3D: Revisiting Change Detection and Captioning from A Video Modeling Perspective (CVPR 2025 · similarity 0.86)
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.