DINO-Mix enhancing visual place recognition with foundational vision model and feature mixing

Area of Science:

• Computer Vision
• Artificial Intelligence
• Geographic Information Systems

Background:

• Visual Place Recognition (VPR) is crucial for determining image locations but struggles in complex environments.
• Current VPR methods show limitations due to lighting variations, seasonal changes, and occlusions.

Purpose of the Study:

• To develop efficient and robust image feature descriptors for high-precision VPR in challenging conditions.
• To enhance the performance of VPR systems in real-world scenarios.

Main Methods:

• Utilized the DINOv2 model for backbone trimming and fine-tuning to extract robust image features.
• Employed a feature mix module to aggregate image features, creating globally robust and generalizable descriptors.

Main Results:

• The proposed DINO-Mix architecture achieved high Top-1 accuracy rates: 91.75% (Tokyo24/7), 80.18% (Nordland), and 82% (SF-XL-Testv1).
• Demonstrated an average accuracy improvement of 5.14% over state-of-the-art (SOTA) methods.
• Experimental results and attention map visualizations confirmed DINO-Mix's superiority in VPR performance.

Conclusions:

• DINO-Mix provides superior VPR performance, especially in complex environments with visual variations.
• The DINO-Mix framework offers a robust and generalizable solution for accurate geographic localization from images.