============================================================ OSINT texture surface re-identification - statistics ============================================================ models present: ['clip', 'clip_l', 'dinov2', 'dinov2_l', 'dinov2_518', 'resnet50', 'efficientnet'] Retrieval (mean [95% bootstrap CI]): CLIP ViT-B/32 mAP=0.535[0.529,0.541] R1=0.943 R@5=0.993 mAP_xs=0.518 CLIP ViT-L/14 mAP=0.590[0.584,0.597] R1=0.954 R@5=0.996 mAP_xs=0.575 DINOv2 ViT-S/14 mAP=0.604[0.598,0.609] R1=0.976 R@5=0.997 mAP_xs=0.586 DINOv2 ViT-L/14 mAP=0.646[0.640,0.651] R1=0.981 R@5=0.997 mAP_xs=0.631 DINOv2 ViT-S/14 (518px) mAP=0.611[0.605,0.617] R1=0.981 R@5=0.999 mAP_xs=0.592 ResNet-50 mAP=0.505[0.499,0.511] R1=0.962 R@5=0.994 mAP_xs=0.484 EfficientNet-B0 mAP=0.552[0.546,0.558] R1=0.970 R@5=0.997 mAP_xs=0.531 Verification (same vs different surface): CLIP ViT-B/32 AUC=0.938[0.935,0.940] EER=0.137 CLIP ViT-L/14 AUC=0.960[0.959,0.962] EER=0.100 DINOv2 ViT-S/14 AUC=0.968[0.966,0.969] EER=0.090 DINOv2 ViT-L/14 AUC=0.973[0.971,0.974] EER=0.072 DINOv2 ViT-S/14 (518px) AUC=0.969[0.967,0.970] EER=0.088 ResNet-50 AUC=0.938[0.936,0.940] EER=0.134 EfficientNet-B0 AUC=0.950[0.948,0.952] EER=0.125 Paired model comparison (Wilcoxon signed-rank on per-query AP): clip vs clip_l: mAP 0.535 vs 0.590 p=3.65e-155 clip vs dinov2: mAP 0.535 vs 0.604 p=9.65e-177 clip vs dinov2_l: mAP 0.535 vs 0.646 p=0.00e+00 clip vs dinov2_518: mAP 0.535 vs 0.611 p=4.43e-211 clip vs resnet50: mAP 0.535 vs 0.505 p=6.95e-35 clip vs efficientnet: mAP 0.535 vs 0.552 p=3.08e-20 clip_l vs dinov2: mAP 0.590 vs 0.604 p=1.96e-08 clip_l vs dinov2_l: mAP 0.590 vs 0.646 p=6.60e-110 clip_l vs dinov2_518: mAP 0.590 vs 0.611 p=5.70e-18 clip_l vs resnet50: mAP 0.590 vs 0.505 p=9.50e-181 clip_l vs efficientnet: mAP 0.590 vs 0.552 p=6.90e-44 dinov2 vs dinov2_l: mAP 0.604 vs 0.646 p=5.78e-128 dinov2 vs dinov2_518: mAP 0.604 vs 0.611 p=1.29e-11 dinov2 vs resnet50: mAP 0.604 vs 0.505 p=0.00e+00 dinov2 vs efficientnet: mAP 0.604 vs 0.552 p=9.99e-144 dinov2_l vs dinov2_518: mAP 0.646 vs 0.611 p=3.72e-70 dinov2_l vs resnet50: mAP 0.646 vs 0.505 p=0.00e+00 dinov2_l vs efficientnet: mAP 0.646 vs 0.552 p=7.56e-289 dinov2_518 vs resnet50: mAP 0.611 vs 0.505 p=0.00e+00 dinov2_518 vs efficientnet: mAP 0.611 vs 0.552 p=2.69e-184 resnet50 vs efficientnet: mAP 0.505 vs 0.552 p=1.23e-154 Best model by mAP: DINOv2 ViT-L/14 mAP by material (best model), worst to best: aluminium_foil 0.489 white_bread 0.504 linen 0.536 cork 0.614 corduroy 0.625 wood 0.639 cotton 0.667 cracker 0.671 brown_bread 0.734 wool 0.786 lettuce_leaf 0.840