Faster or Stronger: Towards Flexible Visual Place Recognition via Weighted Aggregation and Token Pruning

Benchmark Model Rank Results
visual-place-recognition-on-amstertimeWeiAD (DINOv2 ViT-L)–Recall@1: 63.4Recall@5: 83.7
visual-place-recognition-on-mapillary-testWeiAD + WeiToP (DINOv2 ViT-B, retention 0.95)–Recall@1: 76.3Recall@5: 87.9
visual-place-recognition-on-mapillary-valWeiAD (DINOv2 ViT-L)–Recall@1: 92.7
visual-place-recognition-on-mapillary-valWeiAD + WeiToP (DINOv2 ViT-L, retention 0.9)–Recall@5: 96.8
visual-place-recognition-on-pittsburgh-250k-testWeiAD + WeiToP (DINOv2 ViT-L, retention 0.95)–Recall@1: 95.9Recall@5: 99
visual-place-recognition-on-spedWeiAD (DINOv2 ViT-L)–Recall@1: 93.1Recall@5: 97