GSV-Cities

GSV‑Cities Dataset Overview

Dataset Content

• Number of Images: ~530,000
• Number of Locations: >62,000 distinct places
• Geographic Coverage: Multiple cities worldwide
• Image Coverage: Each location has at least 4 images (up to 20)
• Location Spacing: Minimum 100 m between any two locations

Dataset Organization

• Image Naming Convention: city_placeID_year_month_bearing_latitude_longitude_panoid.JPG
• Structure:

├── Images
│   ├── City1
│   │   ├── ...
│   ├── City2
│   │   ├── ...
└── Dataframes
    ├── City1.csv
    ├── City2.csv
    └── ...

• Dataframe Contents: Metadata for each city, convenient for quick Pandas access

Intended Uses

• Performance Boost: Train visual place‑recognition models to achieve state‑of‑the‑art results
• Rapid Training: Enables fast training cycles (≈10‑15 min per epoch)
• Simplified Workflow: No need for offline triplet mining; batches are formed directly
• Rapid Prototyping: Facilitates quick model iteration without lengthy convergence times

Model Evaluation

• Tools: Provided Jupyter Notebook for evaluation
• Metrics: Includes R@1, R@5 across test sets such as Pitts250k‑test, Pitts30k‑test, MSLS‑val, Nordland, etc.
• Pre‑trained Models: ResNet‑50 based models with various output dimensions; performance tables are in the README

Access

• Hosting: Hosted on Kaggle – Kaggle