HoloGeo · Image Geo-localization · Landmark Bias
HoloGeo

Mitigating Landmark Bias in Geo-localization via Evidence-Driven Reasoning

ACM MM 2026

Pengcheng Zhou♠ Xuanyu Liu♦ Yanchen Yin♦ Bobo Li♠ Shengqiong Wu★,* Mong-Li Lee♠ Wynne Hsu♠

♠National University of Singapore, Singapore, Singapore

♦Shandong University of Science and Technology, Qingdao, China

★University of Oxford, Oxford, United Kingdom

Contact:

(* Corresponding author)

News

Quick Overview

What is HoloGeo? HoloGeo systematically studies landmark bias in VLM-based image geo-localization, from diagnosis and measurement to benchmark construction and mitigation. We introduce BI/BH for quantifying landmark bias, LandmarkBias-3K for systematic evaluation, and BF-30K with evidence-driven training to mitigate landmark shortcuts.

Abstract

TL;DR HoloGeo diagnoses landmark bias with BI/BH, builds LandmarkBias-3K and BF-30K, and trains geo-localization VLMs to reason from multiple visual evidence sources rather than landmark shortcuts.

Recent advances in Vision-Language Models have significantly improved image geo-localization, yet existing models remain susceptible to landmark bias, causing them to overlook geographical cues or form spurious correlations. To systematically investigate this issue, we design two quantitative metrics, Bias Intensity and Bias Harmfulness, and establish a comprehensive benchmark, LandmarkBias-3K.

To mitigate landmark bias, we propose HoloGeo, an evidence-driven reasoning framework supported by BF-30K, a high-quality dataset annotated with structured multi-evidence bias-free reasoning chains. With multi-dimensional rewards, HoloGeo encourages balanced attention over diverse visual cues and achieves reliable evidence-driven joint reasoning.

3,000 LandmarkBias-3K diagnostic examples
30K BF-30K bias-free reasoning chains
27.27% City-level accuracy on LandmarkBias-3K
+10.44 City-level gain over Qwen2.5-VL-7B
HoloGeo landmark bias examples
Existing models often over-rely on salient landmark-like cues. HoloGeo integrates complementary geographic evidence to avoid shortcut reasoning.

Insights

The key insight behind HoloGeo is that reliable geo-localization should not collapse onto a single visually salient landmark. Instead, models should reason over multiple evidence types, including architecture, surrounding environment, road signs, landforms, vegetation, and cultural context.

Bias Diagnosis

BI measures how strongly a landmark anchors the prediction, while BH measures whether that influence harms the ground-truth preference.

Bias Benchmark

LandmarkBias-3K contains 3,000 curated examples where landmark cues are insufficient, ambiguous, or deceptive.

Evidence Training

BF-30K provides structured multi-evidence reasoning chains for bias-free geo-localization supervision.

LandmarkBias-3K Construction

Pipeline for constructing LandmarkBias-3K
Images are decomposed into original, landmark-only, and landmark-removed views. BI and BH are computed from model probability shifts to select bias-sensitive examples.

BI/BH Distribution

Empirical distributions of Bias Intensity and Bias Harmfulness
LandmarkBias-3K covers a broad spectrum of harmful landmark influence. Median BI and BH are 2.37 and 0.28.

Global Coverage

Global distribution of benchmark samples
Source images are curated from large-scale geo-localization datasets with broad geographic coverage.

Bias Metric Cases

Representative cases for different BI and BH sign combinations
BI and BH distinguish harmful landmark bias, strong but non-harmful landmark influence, and contextual reinterpretation.

HoloGeo Training Framework

Overview of HoloGeo
HoloGeo first constructs BF-30K with structured multi-evidence reasoning chains, then trains with SFT and GRPO-based rewards for geo-location accuracy, visual grounding, and comprehensive logical reasoning.

Geo-localization Evaluation

HoloGeo improves robustness on the challenging LandmarkBias-3K benchmark while preserving strong performance on standard geo-localization benchmarks.

Standard benchmark accuracy (%) at City / Region / Country levels.
Model IM2GPS IM2GPS3K YFCC4K
GeoReasoner 24.9 / 48.1 / 65.8 26.5 / 40.4 / 57.7 6.3 / 10.4 / 17.3
Geo-R1 42.2 / 59.5 / 77.2 35.4 / 52.3 / 69.7 17.0 / 29.0 / 48.9
GeoChat 38.8 / 59.1 / 76.8 34.7 / 51.7 / 69.4 17.0 / 29.2 / 49.8
GLOBE 44.3 / 59.4 / 76.3 40.2 / 56.2 / 71.5 18.0 / 30.7 / 50.6
HoloGeo 47.3 / 60.3 / 76.8 38.5 / 53.7 / 70.8 18.9 / 31.7 / 51.5
LandmarkBias-3K accuracy (%) at City / Region / Country levels.
Model City Region Country
GPT-4o-mini 19.47 35.77 59.00
Qwen2.5-VL-7B 16.83 28.67 44.57
GeoReasoner 10.67 22.77 47.13
GLOBE 23.27 44.27 63.90
GeoAgent 23.57 45.27 67.07
HoloGeo-SFT+RL 27.27 47.07 68.20
Ablation on IM2GPS with Qwen2.5-VL-7B as backbone.
Model SFT Rgeo Rbox RCLR City Region
Qwen2.5-VL-7B ---- 35.0246.83
Qwen2.5-VL-7B + SFT ✓--- 40.9356.54
HoloGeo Full ✓✓✓✓ 47.2660.34
w/o Rgeo ✓-✓✓ 41.3559.49
w/o Rbox ✓✓-✓ 43.0359.07
w/o RCLR ✓✓✓- 45.9959.49
Landmark type breakdown on LandmarkBias-3K: Qwen2.5-VL-7B vs. HoloGeo.
Type Qwen City HoloGeo City Qwen Region HoloGeo Region Qwen Country HoloGeo Country
Landform8.4218.4830.1650.8241.3068.48
Cultural Cues5.2512.5010.2522.7523.0044.75
Climate13.3313.3320.0026.6733.3346.67
Architecture22.0631.4635.5451.7954.7073.73
Vegetation6.8014.5612.6236.8924.2769.90
Language Signs17.5424.6023.5940.7335.8958.67

Why Landmark Bias Occurs

Attention comparison across geolocation models
Baseline models concentrate attention on salient landmarks. HoloGeo distributes attention across landmark and non-landmark regions for more balanced evidence aggregation.

LandmarkBias vs. IM2GPS

Comparison between LandmarkBias-3K and IM2GPS
LandmarkBias-3K targets misleading salient cues that are not captured by standard benchmark accuracy alone.

Error Analysis

Geo-localization error analysis
Error patterns motivate explicit evidence coverage and balanced visual grounding in HoloGeo.

Case Study

Case study comparing reasoning across models
HoloGeo resists landmark-induced shortcuts by incorporating surrounding architecture, scene setting, and broader geographic context.

Examples

Landmark Bias Examples Benchmark Construction Training Pipeline Reasoning Case Study

BibTeX

Accepted at ACM Multimedia 2026. Citation details will be updated after publication.

@inproceedings{zhou2026hologeo,
  title     = {HoloGeo: Mitigating Landmark Bias in Geo-localization via Evidence-Driven Reasoning},
  author    = {Zhou, Pengcheng and Liu, Xuanyu and Yin, Yanchen and Li, Bobo and Wu, Shengqiong and Lee, Mong-Li and Hsu, Wynne},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia},
  year      = {2026}
}