HoloGeo: Evidence-Driven Vision-Language Model for Image Geo-localization
Mitigating Landmark Bias through Multi-Evidence Geospatial Reasoning
Model Description
HoloGeo is an open-source vision-language model for robust image geo-localization and geospatial reasoning. It is designed to mitigate landmark bias, where a model over-relies on a salient building, statue, text sign, cultural symbol, or other prominent cue while overlooking complementary evidence from the surrounding environment.
Instead of mapping one salient landmark directly to a location, HoloGeo analyzes multiple image regions and integrates evidence from architecture, vegetation, terrain, climate, infrastructure, text, and cultural context before predicting a country and city.
HoloGeo is introduced in HoloGeo: Mitigating Landmark Bias in Geo-localization via Evidence-Driven Reasoning, accepted at ACM Multimedia 2026. It is trained with BF-30K, which contains structured multi-evidence reasoning chains, and evaluated with LandmarkBias-3K, a diagnostic benchmark for misleading-landmark scenarios.
HoloGeo Ecosystem
- HoloGeo: an evidence-driven vision-language model for image geo-localization.
- BF-30K: a 30K training dataset with region-level visual analyses and structured multi-evidence reasoning chains.
- LandmarkBias-3K: a 3K-image diagnostic benchmark for landmark-induced reasoning errors.
- Bias Intensity (BI) and Bias Harmfulness (BH): complementary metrics that measure the strength and harmfulness of landmark influence on model predictions.
Key Features
- Image geo-localization: predicts country- and city-level locations from visual and contextual geographic cues.
- Evidence-driven reasoning: combines information from multiple informative regions instead of relying on a single salient landmark.
- Landmark-bias mitigation: explicitly targets harmful over-reliance on recognizable or visually dominant cues.
- Structured geospatial reasoning: produces region analyses, aggregated reasoning, and a final location using an
<Analyze>-<Think>-<Answer> structure.
- Diagnostic evaluation: evaluated on LandmarkBias-3K and standard image geo-localization benchmarks.
Model Details
| Item | Description |
|---|
| Base model | Qwen/Qwen2.5-VL-7B-Instruct |
| Model type | Vision-language model for image-to-text generation |
| Training | LoRA supervised fine-tuning followed by GRPO reinforcement learning |
| Training data | BF-30K |
| Output | Structured visual analysis, geographic reasoning, country, and city |
| Precision | BF16 |
The released checkpoint contains the merged model weights. The LoRA adapter from the reinforcement-learning checkpoint has been merged into the base model, so the model can be loaded directly with Transformers.
Load
1from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
2
3model_id = "PPKQ/HoloGeo"
4
5model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
6 model_id,
7 torch_dtype="auto",
8 device_map="auto",
9)
10processor = AutoProcessor.from_pretrained(model_id)
Training
HoloGeo uses a two-stage training strategy built on Qwen2.5-VL-7B-Instruct:
- Supervised fine-tuning: LoRA fine-tuning on 1.5K BF-30K samples establishes the structured multi-evidence reasoning format.
- GRPO reinforcement learning: the remaining training samples are used with rewards for output format, geo-localization accuracy, visual-region grounding, and comprehensive logical reasoning.
Training code and example launch scripts are available in the
HoloGeo GitHub repository.
Evaluation
Predictions are evaluated by the geographic distance between the predicted and ground-truth coordinates:
| Granularity | Distance threshold |
|---|
| City | 25 km |
| Region | 200 km |
| Country | 750 km |
Datasets
The accompanying
HoloGeo Dataset provides:
- BF-30K, used for structured multi-evidence geo-localization training;
- LandmarkBias-3K, used to evaluate robustness in ambiguous or misleading landmark scenes.
The dataset has its own license and usage terms. The model's Apache 2.0 license does not override the licenses, copyrights, or terms associated with dataset annotations or upstream images.
Intended Use
HoloGeo is intended for research on image geo-localization, multimodal geographic reasoning, landmark bias, model robustness, and related evaluation methods.
Image geo-localization can create privacy, surveillance, and spatial-security risks. Do not use HoloGeo for unlawful tracking, privacy infringement, identifying private individuals, evading safety measures, criminal activity, or other harmful purposes.
Limitations
- Predictions may be incorrect or overconfident, especially for remote regions, rare landforms, visually similar locations, and areas underrepresented in the training data.
- Generated visual analyses and reasoning may contain hallucinations, factual errors, or unsupported geographic associations.
- Mitigating landmark bias does not eliminate other geographic, cultural, linguistic, or dataset biases.
- Performance depends on image quality, geographic coverage, prompt design, decoding settings, and the availability of informative visual evidence.
- Model outputs should not be treated as independently verified geographic facts.
Citation
If you use HoloGeo, LandmarkBias-3K, or BF-30K in your research, please cite:
1@inproceedings{zhou2026hologeo,
2 title = {HoloGeo: Mitigating Landmark Bias in Geo-localization via Evidence-Driven Reasoning},
3 author = {Zhou, Pengcheng and Liu, Xuanyu and Yin, Yanchen and Li, Bobo and Wu, Shengqiong and Lee, Mong-Li and Hsu, Wynne},
4 booktitle = {Proceedings of the 34th ACM International Conference on Multimedia},
5 year = {2026}
6}
Acknowledgements
HoloGeo builds upon Qwen2.5-VL and uses or draws on resources and tools including MP-16, Google Landmarks Dataset v2, GroundingDINO, InternVL3, DeepSpeed, vLLM, and ModelScope Swift. We thank their authors and open-source communities.
