DINOv3 Feature Adaptation for UAV-to-Satellite Building Retrieval

作者：YiNan Yuan, LiXin Gong, Xin Zhou, Xuerong Yang · 发表于：2026 IEEE 27th China Conference on System Simulation Technology and its Applications (CCSSTA) · 年份：2026 · DOI：10.1109/ccssta69471.2026.11635433

In GNSS-denied or satellite-navigation-degraded environments, unmanned aerial vehicles (UAVs) require visual information to support reliable localization. UAV-to-satellite building instance retrieval provides a feasible solution by matching UAV-view images with a satellite-view gallery, but it remains challenging because of large viewpoint, scale, and appearance variations. We construct a shared-weight two-tower retrieval framework based on DINOv3. The proposed framework extracts patch-level features using DINOv3 and aligns UAV-view and satellite-view representations through a symmetric InfoNCE loss. To investigate how DINOv3 features can be effectively adapted to cross-view retrieval, several feature aggregation strategies are compared, including GAP, GMP, GeM, and MixVPR, and the influence of different backbone unfreezing ranges on retrieval performance and memory consumption is further analyzed. Experiments on the University-1652 dataset show that MixVPR achieves the best performance among the aggregation strategies, with 80.14% rank-1 recall and 86.57% mAP. In the fine-tuning comparison, unfreezing the last 20 Transformer blocks achieves the highest retrieval performance, with 81.76% rank-1 recall and 87.81% mAP. Meanwhile, unfreezing the last 8 Transformer blocks obtains near-optimal performance with lower memory consumption, showing a better trade-off between accuracy and computational cost. These results suggest that DINOv3 adaptation should consider both patch-token a...