Research map / Image generation / Autoregressive and latent generation
GeoBiaset: A Counterfactual Benchmark for Demographic Bias in World-Level Geolocalization
Niccolò Niccoli, Silvia Dani, Marco Mistretta, Andrew Bagdanov, Marco Bertini, Lorenzo Seidenari
Read the paperCheck your idea against 137,217 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 GeoBiaset, a counterfactual benchmark for auditing demographic bias in world-level image geolocalization by examining which visual cues drive regional disparities.
Problem
Current world-level image geolocalization models show biases (e.g., missing cities in training data, better performance in wealthier regions), but it is unclear which visual cues drive these disparities.
Approach
A counterfactual benchmark (GeoBiaset) that is meant to isolate visual cues responsible for geolocalization disparities across demographic/regional groups; the abstract is truncated so the exact construction is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| insert a person into the background image | GeoCLIP · GeoBiaset | continent accuracy (%): 81.9 -> 55.3 | Hurt |
| insert a person into the background image | Img2Loc · GeoBiaset | continent accuracy (%): 70.3 -> 37.5 | Hurt |
| insert a person into the background image | G3 · GeoBiaset | continent accuracy (%): 81.6 -> 47.8 | Hurt |
| insert a person into the background image | GLM-4.1V (zero-shot MLLM) · GeoBiaset | continent accuracy (%): 82.8 -> 69.2 | Hurt |
| mask the inserted person (black) on both original and edited images (ablation) | GeoCLIP · GeoBiaset | continent accuracy (%): 55.3 -> 72.7 | Improved |
| mask the inserted person (gray) on both original and edited images (ablation) | GeoCLIP · GeoBiaset | continent accuracy (%): 55.3 -> 71.7 | Improved |
| blur faces of inserted person (ablation) | GeoCLIP · GeoBiaset | accuracy at 2500km: +8.8 (Black), +11.7 (Latino Hispanic), +1.7 (White) | Improved |
| use race-associated attire for inserted person (ablation) | GeoCLIP · GeoBiaset, Black subject | continent accuracy change (%): -15.5 | Hurt |
| use race-associated attire for inserted person (ablation) | GeoCLIP · GeoBiaset, Middle Eastern subject | continent accuracy change (%): +0.4 | No clear effect |
| scene contains a landmark vs not (ablation) | GeoCLIP · GeoBiaset | POPI-H accuracy change (%): 27.1 -> 9.2 | Improved |
| scene contains a landmark vs not (ablation) | GeoCLIP · GeoBiaset | POPI-D accuracy change (%): 33.2 -> 26.0 | Improved |
| use dedicated geolocalization model instead of zero-shot MLLM | geolocalization models · GeoBiaset | relative continent drop: 33-47% vs 16-25% | Hurt |
Limitations
- Full text unavailable; only the truncated abstract was used.
Experimental setup
Benchmarks: GeoBiaset
Closest papers by meaning
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.88)
- PIGEON: Predicting Image Geolocations (CVPR 2024 · similarity 0.88)
- GeoLLM: Extracting Geospatial Knowledge from Large Language Models (ICLR 2024 · similarity 0.87)
- G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models (NeurIPS 2024 · similarity 0.87)
- No Filter: Cultural and Socioeconomic Diversity in Contrastive Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- Large Language Models are Geographically Biased (ICML 2024 · 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.
