Research map / Time series / Causal and generative time series
CityPulse: Fine-Grained Assessment of Urban Change with Street View Time Series
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.
Curates a multi-city Google Street View time series dataset with image-level change labels and fine-tunes a Siamese DINOv2-based pairwise change detector, showing it beats generic backbones, time-series pairs, and street-view self-supervised pretraining, and correlates with socio-demographic change in Seattle.
Problem
Detecting physical urban change from street-level imagery at scale is hampered by noisy signals (angles, shadows, lighting) and by small, pixel-labeled change detection datasets with limited spatial and temporal coverage.
Approach
Labels street view time series at the image level for urban change points, generates all pairwise combinations with segment-based labels, and trains a Siamese DINOv2 classifier on these pairs, using the temporal series itself as natural augmentation.
Key results
- Fine-tuned DINOv2 reaches 88.85% accuracy and 87.96 F1 on the test set, the best among ResNet101, DINO, and CLIP backbones.
- Training on time-series pairs (88.85% accuracy) outperforms pairwise data with augmentation (85.63%) and unaugmented pairs (85.99%).
- Detected change points correlate with ACS census-tract change (R^2 0.19 for median income, 0.15 for population) while permits show no significant correlation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune DINOv2 (ViT-B/14) as the change-detection backbone | Siamese street-view change detection model · Street-view change detection test set | Accuracy: 88.85 vs. 87.06 | Helped |
| Train only linear layers while freezing DINOv2 | DINOv2 (ViT-B/14) change-detection model · Street-view change detection test set | Accuracy: 88.20 vs. 87.06 | Helped |
| Use time series data instead of randomly sampled pairwise data (ablation) | DINOv2 (ViT-B/14) change-detection model · Street-view change detection test set | Accuracy: 85.99 -> 88.85 | Helped |
| Use time series data instead of randomly sampled pairwise data (ablation) | DINOv2 (ViT-B/14) change-detection model · Street-view change detection test set | F1-Score: 85.67 -> 87.96 | Helped |
| Apply horizontal flip, color jitter, grayscale, and Gaussian blur to pairwise data (ablation) | DINOv2 (ViT-B/14) pairwise-data model · Street-view change detection test set | Accuracy: 85.99 -> 85.63 | Hurt |
| Add StreetMAE pre-training | ViT-B/16 initialized from DINO · Street-view change detection test set | Accuracy: 88.85 -> 78.49 | Hurt |
| Add StreetBYOL pre-training | ViT-B/16 initialized from DINO · Street-view change detection test set | Accuracy: 88.85 -> 86.25 | Hurt |
| Add an unsupervised segmentation head to StreetBYOL (ablation) | StreetBYOL pre-training · Street-view change detection test set | Accuracy: 86.25 -> 87.42 | Helped |
| Add an unsupervised segmentation head to StreetBYOL (ablation) | StreetBYOL pre-training · Street-view change detection test set | F1-Score: 84.78 -> 86.43 | Helped |
Surprising findings
- Domain-specific street-view pretraining (StreetMAE, StreetBYOL) does not surpass generic DINOv2 features.
- Linear probing of frozen DINOv2 outperforms fine-tuning of all other backbones.
Limitations
- Only captures changes visible at street level; interior renovations are missed.
- Google Street View update frequency is inconsistent across regions, especially in developing countries.
- Dataset labels are image-level and limited to change/no-change, with no object-type information.
Experimental setup
Datasets: CityPulse street view time series dataset (Google Street View, 6 cities) Microsoft building footprints TSUNAMI PSCD American Community Survey 5-year estimates Seattle construction permits
Benchmarks: CityPulse street view time series test split
Metrics: accuracy precision recall F1-score linear correlation R^2 p-value
Baselines: ResNet101 DINO CLIP StreetMAE StreetBYOL pairwise training data construction permits
Builds on: DINOv2 BYOL masked autoencoders Siamese networks
Closest papers by meaning
- EMPLACE: Self-Supervised Urban Scene Change Detection (AAAI 2025 · similarity 0.87)
- UrbanFeel：A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective (ICLR 2026 · similarity 0.87)
- UrbanDataLayer: A Unified Data Pipeline for Urban Science (NeurIPS 2024 · similarity 0.85)
- The STVchrono Dataset: Towards Continuous Change Recognition in Time (CVPR 2024 · similarity 0.85)
- CityLens: Evaluating Large Vision-Language Models for Urban Socioeconomic Sensing (ICLR 2026 · similarity 0.85)
- DynamicVL: Benchmarking Multimodal Large Language Models for Dynamic City Understanding (NeurIPS 2025 · similarity 0.84)
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.