GeoGuessr Image Geolocation Model

Humans use subtle cues to guess geographic location: vegetation, soil color, road markings, architecture, sun direction, and even license-plate shape.

Our goal: Can a model do the same from a single image?

• Predict latitude & longitude directly (regression)
• Leverage continent-level priors
• Handle extreme geographic imbalance in OSV-5M

OSV-5M provides nearly five million raw street-view images, but its extreme size makes training difficult. A single 25-epoch run would take multiple days.

To make training feasible, we created a:

• 150,000-image stratified subset preserving the geographic distribution of the full dataset
• Training time reduced from 2+ days → 2.5 hours
• Balanced across 6 continents and 220+ countries
• Increase the size of subsets and test on the increasing subsets step by step

We started with a simple baseline: predict the same (lat, lon) for every image (the dataset mean).

• Latitude mean: 32.94°
• Longitude mean: -11.29°
• Average Haversine distance: 6195.97 km

This baseline lets us evaluate whether any new approach beats a naïve global prior.

We used a pretrained ResNet-18 backbone and added a 5-layer regressor head to predict (latitude, longitude) directly.

However, we learned:

• 1. The world is too large for a single regressor
• 2. Europe and Oceania look visually similar
• 3. South America and South Asia frequently get confused
• 4. Model often predicts locations in the ocean

We experimented with a full country classifier.

Why not Vision Transformers?

• Though ViT performs best with very large, clean datasets, but our time and source constraints limited us
• OSV-5M is noisy and globally imbalanced
• Training ViT was slow and accuracy resembled random guessing

Why ResNet50?

• Stronger for natural imagery
• More efficient to fine-tune
• Better generalization on the subset

Why we didn't use ResNet50 and Country Classifier?

• We got really high training accuracy (98%), but validation accuracy was only 59.83%, which indicates overfitting
• OSV-5M has a large number of images concentrated on a few countries
• Country classifier is more complex and requires more fine-tuning

Our final architecture splits the world into logical regions:

• Stage 1 — Continent Classifier (RouterNet)
• Stage 2 — 6 continent-specific regressors

Each regressor fine-tunes the ResNet-18 backbone on its own regional dataset. This drastically improves accuracy, especially in well-represented regions such as:

• North America
• Central Europe
• Middle East

We deployed the model on HuggingFace Spaces for interactive testing. The model shows strong performance when visual cues align with dataset density. You can try our model here: GeoLocator - Predict latitude/longitude from an image

Strengths

• Predicts with very little context
• Succeeds in regions with strong visual identity (e.g., Middle East)
• Good at densely represented areas of the dataset

Weaknesses

• Cannot understand languages, signs, or landmarks
• Confuses regions with similar vegetation or architecture
• Still predicts ocean locations occasionally

• Big datasets require strategy (sampling, subsets, caching)
• Model choice matters — ViT is not always the answer
• Decompose complex problems (classification → regression)
• Geolocation requires both coarse and fine reasoning

• PIGEON (Haas et al., 2024) — Geolocation using semantic geocells.
• GeoLocSFT (Yi & Shan, 2025) — Fine-tuning multimodal models for geolocation.
• IMAGEO-Bench (Li et al., 2024) — Benchmarking LLMs for spatial reasoning.