NYU-VPR: Long-Term Visual Place Recognition Benchmark with View Direction and Data Anonymization Influences

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Visual place recognition (VPR) is essential for autonomous driving and navigation for the visually impaired.
• Large-scale VPR systems face challenges including varied image view directions and privacy concerns from urban data.
• The impact of these factors on VPR performance is not well understood.

Purpose of the Study:

• To investigate the influence of image view direction and data anonymization on VPR performance.
• To introduce the NYU-VPR dataset for studying these VPR challenges.
• To provide benchmark results for popular VPR algorithms.

Main Methods:

• Creation of the NYU-VPR dataset with over 200,000 images from a 2km×2km area near NYU, captured throughout 2016.
• Evaluation of several popular VPR algorithms on the dataset.
• Analysis of performance variations based on image view direction (front vs. side) and data anonymization.

Main Results:

• Side views present significantly greater challenges for current VPR algorithms compared to front views.
• Data anonymization demonstrated a negligible impact on VPR performance.
• The study provides explanations and in-depth analysis for these observed performance variations.

Conclusions:

• VPR systems need to be optimized for different view directions, with side views requiring further research.
• Privacy-preserving techniques like data anonymization do not significantly hinder VPR accuracy in metropolitan settings.
• The NYU-VPR dataset serves as a valuable resource for advancing VPR research.