Fine-grained Classification on iNaturalist-19
35.33Top-1 Error RateHiE
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| HiECRM=false, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 35.33 | 2.15 | 0.76 | 1.33 | 2.19 | — | |
| HiECRM=true, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 35.42 | 2.14 | 0.75 | 1.23 | 1.79 | — | |
| HiE-SelfCRM=false, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 36.36 | 2.36 | 0.86 | 1.44 | 2.31 | — | |
| HAFCRM=false, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 36.4 | 2.28 | 0.83 | 1.62 | 2.55 | — | |
| Cross-EntropyCRM=false, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 36.44 | 2.39 | 0.87 | 1.97 | 3.25 | — | |
| HiE-SelfCRM=true, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 36.45 | 2.34 | 0.85 | 1.32 | 1.86 | — | |
| HAFCRM=true, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 36.48 | 2.25 | 0.82 | 1.29 | 1.84 | — | |
| Cross-EntropyCRM=true, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 36.51 | 2.33 | 0.85 | 1.32 | 1.86 | — | |
| Chang et al.CRM=false, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 37.23 | 2.28 | 0.85 | 1.75 | 3.02 | — | |
| Chang et al.CRM=true, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 37.31 | 2.24 | 0.84 | 1.3 | 1.84 | — | |
| Cross-Entropy-HCRM=false, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 37.63 | 2.37 | 0.89 | 1.86 | 2.96 | — | |
| Cross-Entropy-HCRM=true, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 37.7 | 2.33 | 0.88 | 1.33 | 1.87 | — | |
| HXECRM=true, alpha=0.1, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 41.47 | 2.38 | 0.99 | 1.41 | 1.93 | — | |
| HXECRM=false, alpha=0.1, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 41.48 | 2.41 | 1 | 1.77 | 2.69 | — | |
| Soft-labelsCRM=false, beta=30, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 41.67 | 2.32 | 0.97 | 1.5 | 2.23 | — | |
| Soft-labelsCRM=true, beta=30, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 41.99 | 2.31 | 0.97 | 1.4 | 1.91 | — | |
| YOLO-v2CRM=false, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 44.37 | 2.42 | 1.08 | 1.9 | 2.87 | — | |
| YOLO-v2CRM=true, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 45.17 | 2.43 | 1.1 | 1.5 | 1.99 | — | |
| HXECRM=false, alpha=0.6, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 45.45 | 2.24 | 1.02 | 1.7 | 2.55 | — | |
| HXECRM=true, alpha=0.6, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 45.6 | 2.21 | 1.01 | 1.4 | 1.86 | — | |
| Barz & DenzlerCRM=false, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 62.63 | 1.99 | 1.24 | 1.49 | 1.97 | — | |
| Soft-labelsCRM=false, beta=4, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 74.7 | 1.82 | 1.36 | 1.49 | 1.96 | — | |
| Barz & DenzlerCRM=true, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 75.61 | 4.62 | 3.49 | 3.66 | 3.84 | — | |
| Soft-labelsCRM=true, beta=4, Backbone=ResNet-18, Pre-trained=ImageNet, Resolution=224x2242023.02 | 77.34 | 2.06 | 1.6 | 1.72 | 2.14 | — | |
| BITS-fixedProtocol=Linear probing, Weight initialization=Frozen ImageNet-1k pre-trained weights2026.02 | — | — | — | — | — | 53.2 | |
| BITS-fixedProtocol=Linear probing, Weight initialization=Self-supervised fine-tuned on target domain2026.02 | — | — | — | — | — | 54.39 | |
| BITS-resetProtocol=Linear probing, Weight initialization=Frozen ImageNet-1k pre-trained weights2026.02 | — | — | — | — | — | 29.74 | |
| BITS-resetProtocol=Linear probing, Weight initialization=Self-supervised fine-tuned on target domain2026.02 | — | — | — | — | — | 49.01 | |
| DINOProtocol=Linear probing, Weight initialization=Frozen ImageNet-1k pre-trained weights2026.02 | — | — | — | — | — | 36.63 | |
| DINOProtocol=Linear probing, Weight initialization=Self-supervised fine-tuned on target domain2026.02 | — | — | — | — | — | 24.82 | |
| SimDINOProtocol=Linear probing, Weight initialization=Frozen ImageNet-1k pre-trained weights2026.02 | — | — | — | — | — | 36.2 | |
| SimDINOProtocol=Linear probing, Weight initialization=Self-supervised fine-tuned on target domain2026.02 | — | — | — | — | — | 29.08 |