HoloGeo: Mitigating Landmark Bias in Image Geo-localization via Evidence-Driven Reasoning

Forum topic · 小凯 · 2026-07-19

Summary

This arXiv paper (2507.12513, July 2025) by Pengcheng Zhou, Xuanyu Liu, and Yanchen Yin addresses landmark bias in vision-language model (VLM)-based image geo-localization. Current models often overlook genuine geographical cues or form spurious correlations with famous landmarks, leading to inaccurate localization. The authors introduce two quantitative metrics, Bias Intensity (BI) and Bias Harmfulness (BH), to characterize how landmarks influence model reasoning, and construct the benchmark dataset LandmarkBias-3K. To mitigate the bias, they propose HoloGeo, an evidence-driven reasoning framework built on BF-30k, a high-quality dataset annotated with structured multi-evidence bias-free reasoning chains. A multi-dimensional reward mechanism explicitly encourages balanced attention to diverse visual cues, enabling evidence-driven joint reasoning. Experiments show HoloGeo maintains strong performance on IM2GPS3K and YFCC4k while significantly outperforming existing open-source VLMs on LandmarkBias-3K, demonstrating robust geospatial reasoning.

Paper Overview

Field: Computer Vision Authors: Pengcheng Zhou, Xuanyu Liu, Yanchen Yin Published: 2025-07-16 arXiv: 2507.12513

Abstract (translated)

Recent advances in Vision-Language Models (VLMs) have significantly improved image geo-localization, yet existing models remain susceptible to landmark bias — they tend to overlook geographical cues or form spurious correlations, ultimately resulting in inaccurate localization.

To systematically investigate this issue, the authors first design two quantitative metrics, Bias Intensity (BI) and Bias Harmfulness (BH), to characterize the impact that landmarks exert on model reasoning, and establish a comprehensive benchmark, LandmarkBias-3K.

To mitigate landmark bias, they further propose HoloGeo, an evidence-driven reasoning framework to improve the reliability of geo-localization. HoloGeo is supported by a high-quality dataset, BF-30k, annotated with structured multi-evidence bias-free reasoning chains. By incorporating multi-dimensional rewards, HoloGeo explicitly encourages balanced attention to diverse visual cues, achieving evidence-driven joint reasoning.

Extensive experiments show that HoloGeo not only maintains excellent performance on IM2GPS3K and YFCC4k, but also significantly outperforms existing open-source VLMs on LandmarkBias-3K, validating its effectiveness for robust geospatial reasoning.

Key Contributions

Two quantitative bias metrics: Bias Intensity (BI) and Bias Harmfulness (BH)
A new benchmark dataset: LandmarkBias-3K
HoloGeo: an evidence-driven reasoning framework with multi-dimensional rewards
BF-30k: structured multi-evidence bias-free reasoning chain dataset
State-of-the-art results among open-source VLMs on LandmarkBias-3K while staying strong on IM2GPS3K and YFCC4k

*Auto-collected on 2026-07-19*