List your company
ResearchFormerly WizWand

Image Classification on NABirds

89.2AccuracyFixRes
4.4426.44548.4570.455Jan 27, 2018Jun 16, 2019Nov 2, 2020Mar 22, 2022Aug 9, 2023Dec 26, 2024May 15, 2026
Updated 5mo ago

Evaluation results

MethodLinks
FixResBackbone=SENet-1542019.0689.2
SENet-154Evaluation Protocol=Baseline2019.0688.3
PAIRS2018.0187.9
CMNBackbone=ResNet-502020.1287.8
Cross-XBackbone=ResNet-502020.1286.2
SSFBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.39, Data Augmentation=Basic2023.1085.9
SSF*Backbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.39, Data Augmentation=Advanced (cutmix, mixup, label smoothing)2023.1085.7
ARC*Backbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.25, Data Augmentation=Advanced (cutmix, mixup, label smoothing)2023.1085.7
LoRABackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.442023.1085.6
ARCBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.252023.1085.3
ARCattBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.222023.1085
HACEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Label Smoothing Condition=No label smooth.2026.0584.96
HACEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Label Smoothing Condition=Soft label, Beta=302026.0584.96
HACEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Label Smoothing Condition=Label smooth.2026.0584.78
HACEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Label Smoothing Condition=Soft label, Beta=102026.0584.7
AdapterBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.412023.1084.3
BiasBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.282023.1084.2
VPT-DeepBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.852023.1084.2
FT-ResNetBackbone=ResNet-50, re-implementation=true2020.1284
SCEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Label Smoothing Condition=No label smooth.2026.0583.5
SCEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Label Smoothing Condition=Soft label, Beta=302026.0583.5
SCEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Label Smoothing Condition=Label smooth.2026.0583.02
MaxEntBackbone=DenseNet-1612020.1283
SCEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Label Smoothing Condition=Soft label, Beta=102026.0583
PC-DenseNet-1612019.0682.8
Full fine-tuningBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=85.982023.1082.7
SupervisedStage 2 pre-training=false, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0680.09
BYOL+CVSAStage 2 pre-training=true, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0679.64
Bilinear CNNReference=PAMI 20172018.0179.4
BilinearCNNBackbone=VGG-162020.1279.4
ResNet-50 BaselineBackbone=ResNet-502018.0179.2
HXEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Alpha=0.22026.0579.09
BYOL+DiLoStage 2 pre-training=true, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0679.04
BYOLStage 2 pre-training=false, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0678.89
InsLocStage 2 pre-training=true, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0678.86
VPT-ShallowBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.252023.1078.8
SimCLRStage 2 pre-training=false, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0676.3
BYOLStage 2 pre-training=true, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0676.12
Linear probingBackbone=ViT-B/16, Pre-trained=ImageNet-21k, Params(M)=0.182023.1075.9
MoCo.v2Stage 2 pre-training=false, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0675.64
MoCo.v2Stage 2 pre-training=true, Evaluation protocol=fine-tuned evaluation, Stage 1 pre-training dataset=ImageNet-1k, Backbone=ResNet-50, Pre-training epochs=8002021.0673.96
SCEArchitecture=ResNet-34, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0571.01
HACEArchitecture=ResNet-34, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0569.44
MaxEntBackbone=ResNet-502020.1269.2
HACEArchitecture=ResNet-18, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0567.93
HACEArchitecture=ResNet-50, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0567.9
R12Student Model=LCNet-35, Teacher Selection Metric=R122026.0567.8
HXEEvaluation Protocol=Linear probing, Backbone=DINOv2-Large, Alpha=0.52026.0564.86
SCEArchitecture=ResNet-18, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0562.94
TACStudent Model=LCNet-35, Teacher Selection Metric=TAC2026.0562.4
SSPStudent Model=LCNet-35, Teacher Selection Metric=SSP2026.0562.4
SCEArchitecture=ResNet-50, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0559.07
HACEArchitecture=ConvNeXt-T, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0550.74
HACEArchitecture=Swin-T, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0544.01
SCEArchitecture=Swin-T, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0541.51
SCEArchitecture=ConvNeXt-T, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0530.35
HACEArchitecture=ViT-B/16, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0529.16
SCEArchitecture=ViT-B/16, Evaluation Protocol=End-to-end training, Aggregation Strategy=Best across all dilution values d in {0.2, 0.5, 0.7} and LR pairings2026.0527.44
CEStudent Model=LCNet-35, Teacher Selection Metric=None (Cross-Entropy)2026.0522.9
PLeaSBackbone=ViT, Model Style=CLIP, Initialization=Same2024.078.3
MuDSCBackbone=ViT, Model Style=CLIP, Initialization=Same2024.078
Simple AvgBackbone=ViT, Model Style=CLIP, Initialization=Same2024.077.8
RegMeanBackbone=ViT, Model Style=CLIP, Initialization=Same2024.077.7