IC-216DINOv2's value (V) facet from self-attention at layer n-1 encodes the most effective local features for VPR re-ranking, outperforming query and key facets, and layer n-1 outperforms the final layer n
Issar Tzachor, Boaz Lerner, Matan Levy, Michael Green, Tal Berkovitz Shalev, Gavriel Habib, Dvir Samuel, Noam Korngut Zailer, Or Shimshi, Nir Darshan, Rami Ben-Ari
The paper ablates which self-attention component (Q, K, V) and which layer of DINOv2 yields the best local patch descriptors for VPR re-ranking. The value facet consistently outperforms query and key on both Tokyo24/7 and MSLs-val. For layer selection, the penultimate layer (n-1) gives the highest recall, while using the final output layer (n) degrades performance, indicating that the most useful local structure is present in the penultimate layer but is lost or altered in the final layer.
Ablations are conducted on two datasets (Tokyo24/7 and MSLs-val) with the re-ranking applied to top-100 candidates; the layer n degradation is modest (1.6 points on Tokyo24/7).