Reproduce “HoloGeo: Mitigating Landmark Bias in Geo-localization via Evidence-Driven Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pengcheng Zhou, Xuanyu Liu, Yanchen Yin, Bobo Li, Shengqiong Wu, Mong-Li Lee, Wynne Hsu
This research from the National University of Singapore and collaborators investigates and mitigates landmark bias in Vision-Language Model (VLM)-based geo-localization. It introduces new metrics (Bias Intensity, Bias Harmfulness) and a diagnostic benchmark (LandmarkBias-3K) to quantify this bias, proposing HoloGeo, an evidence-driven framework that achieved 47.3% city-level accuracy on IM2GPS and significantly improved robustness on LandmarkBias-3K, reaching 27.27% city-level accuracy compared to under 20% for most baselines.
Key insights
- Landmark bias is a significant issue in VLM geo-localization, causing over-confident and erroneous predictions when salient cues are ambiguous or replicated across locations.
- Encouraging balanced, multi-evidence reasoning through structured training data and diverse reward signals (e.g., format, geo-accuracy, visual grounding, logical consistency) effectively mitigates this bias.
- Quantitative metrics like Bias Intensity and Bias Harmfulness are crucial for systematically diagnosing and understanding the specific adverse effects of landmark bias, rather than just relying on aggregate accuracy scores.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.