AmsterTime

Dataset Overview

The AmsterTime Dataset is a specialized benchmark designed for visual place recognition (VPR) tasks under challenging conditions. Specifically, it focuses on matching historical images with modern street views of the same locations in Amsterdam. As a result, it provides a unique dataset for studying domain shifts in computer vision.

Moreover, the dataset includes carefully curated image pairs. Consequently, it enables researchers to develop models capable of recognizing places across time and visual variations.

Data Sources and Collection

The dataset combines images from multiple high-quality sources. For example:

• Historical Images: Sourced from the Amsterdam City Archive
• Modern Images: Collected from Mapillary street view data

Key Features:

• Historical Archive: Over 1,200 license-free images from the Amsterdam City Archive, documenting urban spaces from the past century.
• Crowdsourced Data: Street view images sourced from Mapillary, ensuring contemporary relevance.
• Human Verification: Matches verified by architectural historians and Amsterdam residents to ensure accuracy.
• Diverse Image Properties: Includes variations in time, structure, occlusion, viewpoint, appearance, and illumination.
• Domain Shift: Highlights significant differences between scanned archival and modern street view images, providing a challenging dataset for Visual Place Recognition (VPR) tasks.

Applications and Use Cases

The AmsterTime Dataset can be used in several advanced applications. For instance:

• Visual Place Recognition (VPR): Match locations across time
• Historical Image Analysis: Study urban development
• Computer Vision Research: Benchmark domain adaptation models
• Geolocation Systems: Improve location-based recognition

Conclusion

The AmsterTime Dataset is a powerful benchmark for visual place recognition under domain shift conditions. Overall, it provides diverse and well-validated data for training advanced models. More importantly, it supports research in understanding how environments change over time and how AI systems can adapt to these changes.

This dataset is sourced from Kaggle.