Research map / Label-efficient and robust learning / Self-supervised and contrastive
EMPLACE: Self-Supervised Urban 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.
Introduces AC-1M, a 1.1M-panorama Amsterdam urban change dataset, and EMPLACE, a self-supervised ViT-B/14 (DINOv2-initialised) trained with an adaptive triplet loss on temporal triplets, evaluated via order prediction, discrete change fine-tuning, zero-shot heatmaps, and a housing-price case study.
Problem
Urban scene change detection has relied on small, supervised, bi-temporal datasets with predefined notions of change, limiting scalability across cities and discovery of unspecified changes.
Approach
Learn change features self-supervised from temporally ordered panorama clusters at the same location, using a triplet loss whose margin scales with the time gap between images, plus a cut-and-flip augmentation that respects panorama circularity.
Key results
- On AMS-Trees, EMPLACE (SI-2 pre-training) reaches .855 accuracy and .885 F1 versus .762 accuracy and .732 F1 for DINOv2 fine-tuned.
- On AMS-Buildings, EMPLACE SI-4 zero-shot reaches .781 accuracy, above fine-tuned DINOv2 (.701) and CLIP (.706).
- Small detected changes correlate positively with housing prices (R^2 = .41, p near 0) across 98 neighbourhoods, while large changes show no reliable correlation (R^2 = .09).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with SI-2 triplet intervals instead of no training | ViT-B/14 with DINOv2 features · Order prediction, all test sets | accuracy: .616 -> .936 | Helped |
| Use SI-1 triplet intervals rather than SI-2 | EMPLACE · Order prediction, all test sets | accuracy: .936 -> .744 | Hurt |
| Use SI-3 triplet intervals rather than SI-2 | EMPLACE · Order prediction, all test sets | accuracy: .936 -> .900 | Hurt |
| Use SI-4 triplet intervals rather than SI-2 | EMPLACE · Order prediction, all test sets | accuracy: .936 -> .840 | Hurt |
| Fine-tune EMPLACE features rather than DINOv2 features | Siamese discrete-change classifier · AMS-Buildings | accuracy: $.701\pm.023$ -> $.732\pm.031$ | Helped |
| Use zero-shot EMPLACE SI-4 rather than DINOv2 | Discrete-change prediction · AMS-Buildings | accuracy: $.701\pm.023$ -> $\textbf{.781}\pm.022$ | Helped |
| Use fine-tuned EMPLACE rather than DINOv2 | Siamese discrete-change classifier · AMS-Trees | accuracy: $.762\pm.028$ -> $\textbf{.855}\pm.003$ | Helped |
| Use fine-tuned EMPLACE rather than DINOv2 | Siamese discrete-change classifier · AMS-Trees | F1: $.732\pm.051$ -> $\textbf{.885}\pm.001$ | Helped |
| Use zero-shot SI-4 rather than zero-shot SI-2 | EMPLACE · AMS-Buildings | accuracy: $.761\pm.025$ -> $\textbf{.781}\pm.022$ | Helped |
Surprising findings
- SI-1 EMPLACE wins on its own test set but does not generalise to other test sets, which the authors attribute to an overly easy same-season training objective.
- EMPLACE SI-4 zero-shot outperforms SI-2 zero-shot on AMS-Buildings accuracy and precision.
Limitations
- Large-scale change labels are not available, so discrete change evaluation relies on newly constructed small datasets (AMS-Buildings, AMS-Trees).
- Only Amsterdam is studied; generalisation to other cities is not shown.
- Large-change correlation with housing prices is not statistically significant (R^2 = .09).
Experimental setup
Datasets: AC-1M AMS-Buildings AMS-Trees PSCD CityPulse TSUNAMI Amsterdam housing prices
Benchmarks: AC-1M order prediction AMS-Buildings AMS-Trees
Metrics: accuracy precision recall F1
Baselines: ResNet101 CLIP DINOv2
Builds on: DINOv2 triplet loss DBSCAN
Closest papers by meaning
- Towards Generalizable Scene Change Detection (CVPR 2025 · similarity 0.88)
- CityPulse: Fine-Grained Assessment of Urban Change with Street View Time Series (AAAI 2024 · similarity 0.87)
- The STVchrono Dataset: Towards Continuous Change Recognition in Time (CVPR 2024 · similarity 0.84)
- MUCD: Unsupervised Point Cloud Change Detection via Masked Consistency (AAAI 2025 · similarity 0.84)
- Change3D: Revisiting Change Detection and Captioning from A Video Modeling Perspective (CVPR 2025 · similarity 0.84)
- UrbanFeel：A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective (ICLR 2026 · similarity 0.83)
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.