HoloGeo Dataset

Training Data and a Diagnostic Benchmark for Landmark-Bias Research in Image Geo-localization

🌐 Project Page · 📄 Paper · 💻 Code · 🤖 Model

---

📌 Dataset Overview

HoloGeo Dataset provides two core data resources for studying landmark bias in image geo-localization:

• BF-30K: structured multi-evidence training data;
• LandmarkBias-3K: a benchmark for diagnosing landmark-induced reasoning errors.

In image geo-localization, vision-language models may over-rely on highly salient cues such as prominent buildings, statues, textual signs, or cultural symbols, while overlooking complementary geographic evidence from the overall scene, resulting in unreliable localization. We refer to this phenomenon as landmark bias.

HoloGeo Dataset studies this problem from both training and evaluation perspectives:

Dataset	Primary purpose	Scale
BF-30K	Trains region-level, structured, multi-evidence geographic reasoning	30k samples
LandmarkBias-3K	Diagnoses model reasoning robustness in ambiguous or misleading landmark scenes	3k test images

This repository releases data resources only. Model weights are available at PPKQ/HoloGeo, and training and inference code is available at PKQZPC/HoloGeo.

---

📊 Landmark Bias Metrics

HoloGeo introduces two complementary metrics to quantify the impact of landmarks on model predictions and to construct and analyze LandmarkBias-3K:

• Bias Intensity (BI): measures the anchoring strength of a landmark on landmark-induced predictions;
• Bias Harmfulness (BH): measures whether a landmark weakens the model's relative preference for the ground-truth location.

BI characterizes the strength of landmark influence, while BH further determines whether this influence harms correct localization.

---

📁 Repository Contents

HoloGeo/
└── data/
    ├── BF-30K/
    │   └── BF-30K.jsonl
    └── LandmarkBias-3K/
        ├── Test_dataset.jsonl
        └── Test_dataset_images.tar.zst

Note: Test_dataset_images.tar.zst is the accompanying image archive for LandmarkBias-3K. It is not loaded as a dataset split by load_dataset or the Dataset Viewer.

---

🚀 Loading the Dataset

Install Hugging Face Datasets:

pip install datasets

Load the two data configurations separately:

from datasets import load_dataset

bf30k = load_dataset(
    "PPKQ/HoloGeo",
    "BF-30K",
    split="train",
)

landmark_bias = load_dataset(
    "PPKQ/HoloGeo",
    "LandmarkBias-3K",
    split="test",
)

For image-based LandmarkBias-3K evaluation, also download and extract:

data/LandmarkBias-3K/Test_dataset_images.tar.zst

In a Linux/macOS environment with zstd installed (which provides unzstd), extract the archive with:

tar --use-compress-program=unzstd \
  -xf data/LandmarkBias-3K/Test_dataset_images.tar.zst

The image field in Test_dataset.jsonl corresponds to the filename of the extracted image.

---

🧩 BF-30K

BF-30K trains models to perform image geo-localization through region-level, multi-evidence reasoning, and contains 30k samples.

Fields

Field	Description
image_id	Image identifier in the source dataset
url	Original image URL
IMG_ID	Source image identifier retained for data traceability
latitude, longitude	Geographic coordinates associated with the image
original_dataset	Source dataset identifier
Box	List of geographically discriminative regions
Box[].Box	Bounding-box coordinates for one region
Box[].Analysis	Analysis of the region's visual content and geographic relevance
vlm_think	Structured reasoning that aggregates evidence from multiple regions
vlm_answer.country	Country-level supervision target
vlm_answer.city	City-level supervision target

---

🧭 LandmarkBias-3K

LandmarkBias-3K is a diagnostic benchmark for landmark bias and contains 3k test images.

The benchmark primarily covers the following challenging cases:

• A landmark is visually prominent but insufficient for accurate localization;
• Similar landmarks may appear in different countries or cities;
• Textual, architectural, or cultural cues provide only coarse-grained geographic information;
• A salient cue may mislead the model toward an incorrect location.

Fields

Field	Description
data_source	Source dataset of the sample
image	Filename in the LandmarkBias-3K image archive
country, city	Ground-truth country and city labels
id	Unique sample identifier
lat, lon	Ground-truth geographic coordinates
label	Landmark-bias category

Evaluation Protocol

Accuracy is calculated from the geographic distance between the predicted location and the ground-truth coordinates:

Evaluation granularity	Distance threshold
City	<= 25 km
Region	<= 200 km
Country	<= 750 km

---

⚠️ Data Use, Limitations, and Responsible Use

• BF-30K references original images only through URLs and does not redistribute image binaries. Upstream images remain subject to their original licenses, copyrights, and terms of use.
• LandmarkBias-3K provides images for research reproduction and benchmark evaluation. Users remain responsible for complying with the relevant licenses and terms of the original data sources.
• Model-generated analyses and reasoning are intended only for research and training, and should not be treated as independently verified geographic facts.
• Do not use this dataset or its derived models for unlawful tracking, privacy infringement, evasion of safety measures, criminal activity, or other harmful purposes.

---

⚖️ License and Third-Party Materials

HoloGeo-provided annotations, reasoning texts, coordinate and metadata organization, and documentation are released under the CC BY-NC 4.0 license.

• The license permits sharing and adaptation for non-commercial purposes. When using this dataset or its derivatives, provide attribution and cite the HoloGeo paper.
• BF-30K provides original image URLs only and grants no copyright, license, or redistribution right for the original images.
• LandmarkBias-3K images and their source metadata remain subject to the original licenses, copyrights, and terms of MP-16, GLDv2, and other upstream sources. This repository does not grant rights beyond those upstream terms.
• Users are responsible for ensuring that their use complies with applicable law, institutional policies, and all upstream-source terms.

---

📖 Citation

@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 ACM International Conference on Multimedia},
  year      = {2026}
}

---

🙏 Acknowledgements

We thank the authors and open-source communities behind MP-16, Google Landmarks Dataset v2, GroundingDINO, Qwen2.5-VL, InternVL3, DeepSpeed, vLLM, and ModelScope Swift.