Weather-Disentangled Scene Retrieval via Contrastive Domain Adaptation in DINOv2

Robust scene retrieval under varying environmental conditions presents a significant challenge for computer vision systems deployed in autonomous navigation and long-term visual localization applications. While foundation models like DINOv2 demonstrate exceptional object recognition capabilities through their self-supervised learning paradigm, our analysis reveals their inherent limitations in scene feature memorization despite superior object-level comprehension. Our proposed solution integrates three key innovations: 1) Synthetic dataset generation via GAN-based weather translation that preserves scene topology, 2) A novel contrastive learning framework where positive pairs are defined as identical scenes under distinct meteorological conditions, and 3) A hybrid adaptation strategy for DINOv2 that maintains foundational visual knowledge through frozen transformer blocks while specializing in high-level scene characterization via fine-tuned terminal layers. Experimental validation on the transformed Cityscapes dataset demonstrated significant improvements in both cross-weather scenario accuracy (achieving 93.5% enhancement) and adjacent frame similarity (20.9% improvement) for the fine-tuned model, particularly under specific weather conditions. Github code: https://github.com/ZihaoChenz/Dinov2_Matching