Long-Term Visual Localization Revisited.

Area of Science:

• Computer Vision
• Robotics
• Geographic Information Systems

Background:

• Visual localization is crucial for autonomous vehicles and augmented reality, requiring robustness to diverse environmental conditions.
• Existing datasets often lack comprehensive ground truth pose information, hindering accurate evaluation of localization methods.
• Variations in lighting, weather, and seasons significantly challenge the performance of six degree-of-freedom (6DOF) camera pose estimation.

Purpose of the Study:

• To extend existing visual localization datasets with accurate ground truth pose information.
• To facilitate the evaluation of visual localization techniques under challenging, real-world viewing conditions.
• To stimulate research in long-term visual localization and learned local image features.

Main Methods:

• Augmenting three publicly available datasets with precise 6DOF camera pose data.
• Establishing a private test set from the extended datasets for unbiased benchmarking.
• Analyzing the performance of state-of-the-art visual localization approaches on the enhanced datasets.

Main Results:

• The extended datasets enable detailed analysis of factors impacting 6DOF camera pose estimation accuracy.
• State-of-the-art localization methods were evaluated, with results largely informed by server submissions.
• A significant portion of ground truth poses were released, with a held-out test set for future research.

Conclusions:

• The enhanced datasets provide a valuable resource for advancing robust visual localization.
• The benchmarking server and released data will foster progress in long-term visual localization and related fields.
• Accurate pose estimation under varied conditions remains a critical area for autonomous systems and AR.