Reference Pose Generation for Long-term Visual Localization via Learned Features and View Synthesis.

Area of Science:

• Computer Vision
• Robotics

Background:

• Visual localization is crucial for autonomous driving and augmented reality.
• Accurate 6 Degree-of-Freedom (DoF) reference poses are essential for benchmarking visual localization methods.
• Traditional Structure-from-Motion (SfM) methods struggle with varying image conditions and manual annotation is not scalable.

Purpose of the Study:

• To develop a semi-automated approach for generating accurate reference poses for visual localization.
• To improve the quality of existing datasets, particularly for nighttime conditions.
• To enhance the performance evaluation of visual localization algorithms.

Main Methods:

• Propose a semi-automated pose generation method using learned features.
• Match features between 3D model renderings and real images.
• Iteratively refine pose estimates based on feature matches against renderings.

Main Results:

• Significantly improved nighttime reference poses for the Aachen Day-Night dataset.
• Demonstrated up to 47% performance improvement for state-of-the-art visual localization methods.
• Extended the dataset with new nighttime images and provided uncertainty estimates for poses.

Conclusions:

• The proposed method offers a scalable and accurate solution for generating high-quality reference poses.
• Improved pose accuracy leads to better benchmarking and development of visual localization systems.
• The enhanced dataset and evaluation criteria will advance research in autonomous driving and augmented reality.