Research map / Vision-language and multimodal / Cross-modal learning
UrbanFusion: Stochastic Multimodal Fusion for Contrastive Learning of Robust Spatial Representations
Dominik J. Mühlematter, Lin Che, Ye Hong, Martin Raubal, Nina Wiedemann
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train all models on the same PP2-M dataset, use UrbanFusion instead of GAIR or GeoCLIP | coordinate-only location encoding · 8 downstream datasets, linear probe | R2 / F1: best on 5 out of 8 datasets | Improved |
| use UrbanFusion instead of GPS2Vec | coordinate-only encoding · Housing Prices and Energy Consumption | R2 | Hurt |
| add multimodal inputs with fused embedding instead of concatenating encoders | UrbanFusion multimodal encoding · 6 downstream tasks, linear probe | R2 / F1: best on 4 of 6 tasks; lower on land cover and coarse land use | Mixed |
| concatenate encoder outputs instead of fusing them (ablation) | UrbanFusion · Land cover and coarse land use | F1 | Improved |
| use multimodal non-coordinate inputs on unseen cities | UrbanFusion · Cross-regional generalization, 6 tasks | R2 / F1: ranks first on 5 out of 6 tasks | Improved |
| use MLP downstream learner instead of linear probe (ablation) | ZIP code-level tasks · ZIP Code tasks | R2 | Hurt |
| add latent modality reconstruction loss alongside contrastive loss with random masking | UrbanFusion · synthetic PID data | capture of redundant, unique and synergistic information | Improved |
| train with only coordinates plus one modality per location (Bimodal) (ablation) | UrbanFusion pretraining · all evaluated domains | relative performance: retains 99.35% of full-pair performance | No clear effect |
| train with partially missing modalities (Partial) (ablation) | UrbanFusion pretraining · ZIP Code multimodal | R2 (%): 75.1 -> 72.3 | Hurt |
Closest papers by meaning
- Diffusion Transformers as Open-World Spatiotemporal Foundation Models (NeurIPS 2025 · similarity 0.86)
- Urban Region Embedding via Multi-View Contrastive Prediction (AAAI 2024 · similarity 0.86)
- Improving Region Representation Learning from Urban Imagery with Noisy Long-Caption Supervision (AAAI 2026 · similarity 0.85)
- UrbanFeel:A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective (ICLR 2026 · similarity 0.85)
- UrBench: A Comprehensive Benchmark for Evaluating Large Multimodal Models in Multi-View Urban Scenarios (AAAI 2025 · similarity 0.84)
- Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach (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.
