Visual Place Recognition Based on an Adaptive D-Value Optimization Strategy

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Visual Place Recognition (VPR) methods like EigenPlaces use Singular Value Decomposition (SVD) based focal points for training.
• A fixed focal distance (D) in VPR training is suboptimal due to varying urban scene geometries.
• The optimal focal distance (D) is dataset-dependent, impacting VPR performance.

Purpose of the Study:

• To analyze the sensitivity of the fixed focal distance (D) in EigenPlaces across diverse datasets.
• To develop an adaptive focal distance strategy that improves VPR performance in varied environments.
• To enhance the environmental perception reliability of intelligent sensing platforms.

Main Methods:

• Systematic analysis of focal distance (D) sensitivity across multiple benchmark datasets.
• Proposal of a depth-aware adaptive D strategy using monocular depth estimation for per-cell focal distances.
• Integration of quantile mapping to ensure variance in assigned D values and principled connection between visual data and geometric supervision.

Main Results:

• The optimal focal distance (D) is confirmed to be highly dataset-dependent, with performance variations up to 4.4%.
• The proposed depth-aware adaptive D strategy achieved the best same-distribution performance on Pitts30k, AmsterTime, and SF-XL benchmarks.
• Ablation studies provided practical guidance for adaptive focal distance selection in VPR training.

Conclusions:

• Adaptive focal distance selection is crucial for robust VPR performance across different urban environments.
• Depth-aware strategies offer a principled way to connect visual data with geometric training supervision for VPR.
• The developed method enhances the reliability and adaptability of VPR systems for intelligent platforms.